Contents, 16 chapters
The 2026 landscape
Search did not get replaced in 2026. It got a layer added on top of it, and that layer decides what most people ever see. The useful question is no longer whether SEO still works. It is which of the two outcomes you are measuring, because they have stopped agreeing with each other.
Up from 6.5% in January 2025. Coverage ranges from 4% to 88% by vertical.
Down from roughly 76% in mid-2024, depending on measurement method.
ChatGPT sits at 17.9%. Google holds roughly 90% of transactional intent.
Fewer visits, arriving much further down the funnel than organic average.
Rising to 77.2% on mobile. Zero-click is the default case now, not the exception.
AI crawlers account for roughly 26.7% of that, split between training and answering.
The one change that reorganizes everything else
In mid-2024, about 76% of AI Overview citations came from pages already ranking in Google's organic top 10. By February 2026 that figure sits somewhere between 17% and 38% depending on whose methodology you accept. Take the widest credible reading and between 62% and 83% of citations now come from pages that do not rank on page one for the query that surfaced them.
That single statistic is why this edition is structured differently from the 2025 one. A page-one report and an AI visibility report used to be roughly the same document. They are not any more, and a team that only runs the first one is flying with half an instrument panel.
What that looks like in a traffic report
The most common diagnostic mistake of 2026 is comparing this year's organic sessions against last year's, seeing a drop, and concluding there is a content quality problem. For a large share of sites there is not. Across publishers, total impressions rose roughly 49% since AI Overviews launched while click-throughs fell about 30%. Positions held. An answer moved in above the results and absorbed the click.
Behavioral research puts the click rate at 8% on queries with an AI Overview against 15% without, and found only about 1% of users clicked a link inside the overview itself. Meanwhile 26% of sessions ended entirely after reading one, against 16% without.
Before you rewrite anything
Segment your keyword set by whether an AI Overview is present, then compare the two cohorts against each other rather than against last year. If the no-overview cohort is stable and the overview cohort is down, you have a visibility problem, not a quality problem, and rewriting healthy pages will not fix it.
The traffic that still arrives is worth more
This is the part that gets lost in the doom framing. AI referral sessions convert at materially higher rates: 15.9% from ChatGPT, 10.5% from Perplexity and 5.0% from Claude, against a Google organic average near 1.76%. Independent analysis puts the aggregate advantage at roughly 4.4x.
Retail data for Q1 2026 shows AI-referred traffic up 393% year over year with 42% better conversion, 37% more revenue per visit and 48% more time on site. Someone who arrives after an assistant has already answered their comparison questions is a different visitor from someone who clicked the fourth blue link.
And citation carries a halo beyond its own click. Brands cited in AI Overviews earn roughly 120% more organic clicks per impression than uncited brands on the same queries, with paid CTR uplift near 91%. Presence in the answer functions as an endorsement that lifts every other listing on the page.
Where the queries actually are
Measured across total digital queries rather than search-engine queries, Google sits near 57% and ChatGPT near 17.9%, with other assistants and engines taking the remainder. Google has roughly 5.3 billion monthly users against ChatGPT's 1 billion, though ChatGPT sessions run more than twice as long.
The split is by intent, not loyalty. Google handles roughly 90% of transactional queries. ChatGPT takes about 64% of generative and creative ones. People ask assistants to think with them and ask Google to buy things. Within the assistant category itself the field is moving fast: ChatGPT's share of assistant web traffic fell from 86.7% to 51.3% between Q2 2025 and Q2 2026 while Gemini rose from 5.7% to 27.7%.
Plan for a multi-surface landscape, not a single successor to Google. Anyone who told you in 2025 that ChatGPT was about to replace search was wrong, and anyone telling you now that the assistant market has settled is making the same error in the other direction.
Signal weights
These are estimated relative influence weights, normalized to 100 and aggregated from correlation studies and practitioner surveys. Google publishes none of this. Treat the direction of travel as reliable and any single percentage as an argument rather than a fact.
Estimated relative influence by category, 2026
Authority
- Content quality and relevance24%
- Authority: links, citations, mentions26%
- Brand and entity signals14%
- E-E-A-T aligned signals13%
- User engagement and satisfaction12%
- Technical and page experience11%
Source. Aggregated from correlation studies and practitioner surveys. Normalized to 100%. Directional, not published by Google.
What moved between editions
The 2025 edition had ten factors. This one has twelve, because two things that used to be folded into other categories now behave differently enough to need their own row: brand and entity signals, and freshness measured against citation rather than ranking.
Reading the table
Brand and entity signals are the story
From roughly 2% to 8% in a single year, and that understates it. Unlinked brand mentions correlate with AI visibility at 0.66 to 0.71 against 0.22 for raw backlink count. Brands in the top quartile by mention volume receive on the order of 10x more AI citations than the rest of the field. This is the one line in the table that should change a budget.
Backlinks became a threshold, not a slope
Analysis of 1,000 domains found the relationship between authority and AI mentions far stronger by rank correlation (Spearman 0.57) than by linear correlation (Pearson 0.23 against AI visibility). That gap is the statistical signature of a threshold: you need to clear an authority floor to be considered at all, after which additional links buy comparatively little.
The same study found follow and nofollow links correlate almost identically with AI visibility (0.334 and 0.340 by Pearson), and image links slightly outperform text links. Read together, that suggests AI systems treat links as evidence that a brand is discussed rather than as votes that accumulate.
Authority did not shrink, it redistributed
Backlinks as an isolated factor sit at 11%. But the Authority category as a whole, meaning links plus citations plus mentions plus trust signals, is still the largest single block of influence at 26%. The change is not that authority stopped mattering. It is that the cheapest way to buy it stopped working.
Freshness hardened into a maintenance requirement
The jump from under 1% to 6% was the 2025 headline. In 2026 it stopped being an opportunity and became an obligation. Pages updated within 30 days earn 3.2x the citations of older equivalents, and pages untouched for three months or more are 3x more likely to lose citations they already held. Roughly 65% of AI Overview citations come from content under a year old and 89% from content under three.
The practical consequence is a change in how content teams should spend their week. Refreshing your twenty best pages now has a higher expected return than publishing three new ones, and most editorial calendars have not caught up to that.
Exact-match keywords keep losing precision
Title tag weight fell from 14% to 11%. Query fan-out is why. When a single prompt is decomposed into a dozen sub-queries you never see, optimizing a title for the string the user typed is optimizing for one input out of many. Cover the intent and its adjacent questions instead.
Technical is still hygiene, with a new failure mode
Mobile-friendliness, HTTPS and fast loading remain baseline expectations where failing costs more than excelling gains. What is new is that client-side rendering has become a way to be invisible. Several AI crawlers execute little or no JavaScript, so content that only exists after hydration may never be retrieved at all. That is a 2026 problem that did not exist in the 2024 version of this list.
Correlation data
Correlation coefficients between factor scores and organic ranking position, aggregated across studies covering more than a million URLs. These describe rankings. They do not describe AI citations, and conflating the two is the most expensive reporting error available in 2026.
Factor correlation with organic ranking position
These measure rankings, not AI citations. The two are no longer the same question.
Scale 0 to 1.00
Source. Aggregated from published industry correlation studies covering 1M+ URLs, data window January 2025 to January 2026, 95% confidence interval. Pearson coefficients between standardized factor scores and SERP position.
How to read these honestly
Strong, 0.80 and above
Content quality (0.89) and E-E-A-T aligned signals in YMYL (0.85) remain the most reliable relationships in the dataset, and they have been stable across four editions of this report. Intent match (0.82) completes the set. Nothing here is surprising, which is itself the point: the top of this list barely moves year to year.
Moderate to strong, 0.65 to 0.80
Topical authority (0.74), backlink quality (0.72) and, new this year, brand and entity signals (0.72). Freshness (0.70) and passage-level structure (0.69) both climbed. Core Web Vitals (0.67) continues to behave as a threshold rather than a lever.
Moderate, below 0.65
Server-rendered content (0.61) is a new entry and reads as a proxy for retrievability. CTR (0.58) stays ambiguous: it is unclear whether higher CTR improves rankings or better rankings produce higher CTR, and Google has said it does not use short-term CTR spikes directly. Schema (0.58) is cheap enough that the ambiguity does not matter.
Method
Pearson coefficients between standardized factor scores and SERP positions 1 to 100, controlled for industry, competition level, geography and device. Data window January 2025 to January 2026, 95% confidence interval, primarily English-language Western market results.
These measure rankings, not AI citations
Every coefficient above describes the relationship between a factor and a position in Google's organic results. As of 2026 that is a different question from what gets cited in an AI answer, and the gap is large: between 62% and 83% of AI Overview citations come from pages outside the organic top 10. The equivalent ranking for AI visibility is in chapter 05, where brand mentions correlate at 0.66 to 0.71 against 0.22 for raw backlinks. Do not read one table as a proxy for the other.
Five caveats that apply to every number on this page
- Correlation is not causation. A high coefficient does not prove the factor causes the ranking. Sites that do one thing well tend to do several things well, and no observational study untangles that.
- Rankings are multi-factor. Nothing here operates alone. A page does not rank because of its freshness score; it ranks because of a combination in which freshness participates.
- Industry variation is enormous. E-commerce sites show higher Core Web Vitals correlation than the aggregate. B2B shows stronger correlation with depth and thought leadership. The per-vertical view is in chapter 10.
- Correlation strength drifts. These are a snapshot of a twelve-month window during which Google shipped four core updates. A coefficient measured in January is not a constant.
- Sample skew is real. Studies concentrate on English-language, Western market, commercially valuable queries, because that is where the tooling points. Treat findings as least reliable exactly where your market is least like that.
The AI search layer
Traditional SEO optimizes for one event: a position in a ranked list. An AI answer is assembled in four separate stages, and a page can fail at any one of them while looking perfectly healthy in the others. Almost every credible finding in the generative engine optimization literature is conditional on which stage it applies to, and most of the bad advice comes from ignoring that.
The four-stage answer pipeline
One prompt is decomposed into many concurrent sub-queries covering subtopics, comparisons and edge cases. You are not competing for the query the user typed. You are competing for a dozen synthetic queries you never see.
Each sub-query pulls a candidate pool from the index, from grounding sources and from the model's own priors. This is the stage most GEO advice ignores, and the stage that decides whether anything else you do matters at all.
The model reranks the pool and picks passages to ground each claim. Position inside the retrieved context matters more here than most content rewrites do. Extractable, attributed statements win at this stage.
A subset of the passages actually used receives a visible citation. Brands are routinely mentioned without being cited and cited without being mentioned. On Gemini the overlap between the two can be as low as 30%.
Why the stage matters more than the tactic
The widely repeated "40% visibility gain" from the original GEO research was measured on documents that had already been retrieved. It describes stage three. End-to-end testing that includes retrieval and reranking found body-only rewrites can reduce top-20 presence by roughly 9% and top-10 presence after reranking by roughly 16%. Optimizing for the generator while ignoring retrieval can make a page harder to find, not easier.
AI Overview coverage by vertical
Share of queries returning an AI Overview, 2025 against February 2026
View the underlying numbers
| Vertical | 2025 | February 2026 |
|---|---|---|
| Healthcare | 72% | 88% |
| Education | 18% | 83% |
| B2B technology | 36% | 82% |
| Restaurants | 10% | 78% |
| Insurance | 24% | 63% |
| All US queries | 6.5% | 48% |
| E-commerce | 29% | 4% |
Source. BrightEdge, February 2026. Prevalence varies with keyword set: studies using short head terms report materially lower coverage, so treat 20% to 50% as the defensible range by vertical.
Coverage expanded, then diverged
AI Overviews reached roughly 48% of US queries by February 2026, up from 6.49% in January 2025. But the aggregate hides the interesting part. Education went from 18% to 83%. Restaurants went from 10% to 78%. B2B technology from 36% to 82%.
E-commerce went the other way, from 29% down to about 4%, as Google routed commercial intent toward Shopping surfaces and ad units instead of generative answers. If you sell things, the AI Overview is largely not your problem. If you explain things, it is the whole problem.
Prevalence numbers disagree for real reasons. Studies report AI Overview coverage anywhere from 16% to 50% depending on keyword set, because short head terms trigger far less often than long informational queries. A defensible range is 20% to 50% by vertical. Anyone quoting a single global number without naming their keyword methodology is quoting a marketing figure.
AI Mode is a different product, not a bigger overview
Average words per AI Mode query, against 4.0 for traditional search and 23 or more for ChatGPT prompts
Queries per AI Mode session, against 5 or more in a traditional Google session
Of AI Mode searches that end without a click to any external domain
Of AI Mode sessions that generate a click out to a website at all
Content written for a four-word head term is structurally mismatched to a conversational, comparison-shaped prompt. The register is different, the expected answer length is different, and the sub-questions the system fans out to are different. This is the single clearest argument for writing comparison and consideration content rather than more keyword-targeted pages.
Clicks fell, then partially recovered
Longitudinal tracking across 2.43 billion impressions found click-through on AI Overview queries falling from 1.76% in June 2024 to a floor of 0.61% in September 2025, then rebounding to 2.4% by February 2026 as Google reworked link placement. Non-AIO queries still convert attention better at 3.8%.
Position-one CTR impact estimates range from -15.5% in one 700,000-keyword study to -34.5% in a 300,000-keyword study, widening to -37% when a featured snippet is also present. The range is wide because the methodologies differ; the direction is not in dispute.
Read every AI search statistic with these five caveats
- Answers are not deterministic. Repeated runs of the same prompt change citation selection by 9% to 28% within 24 hours, even at temperature zero. Single-query checks are noise.
- Correlation is doing heavy lifting. Nearly all published AI visibility factors are correlational. Mention volume and citation rate rise together, which does not establish that buying mentions produces citations.
- Lab gains do not survive competition. Controlled benchmarks show individual GEO gains eroding toward zero as adoption spreads. Early movers capture a redistribution, not new visibility.
- Vendor data has a direction. Most large citation datasets are published by companies selling AI visibility tooling. Directionally useful, worth discounting on magnitude.
- Denominators matter. "Cited in 40% of prompts" means nothing without knowing how many of those prompts triggered an AI answer at all.
What drives AI citations
Ranked by measured strength rather than by how often the tactic gets recommended. The ordering below is the most useful thing in this report, because it inverts the priority list most SEO teams are still working from.
What correlates with AI search visibility
Spearman correlation coefficients across a 75,000-brand dataset
Scale 0 to 0.800
Source. Ahrefs, 75,000 brands, brands with Domain Rating above 40 and keywords above 800 monthly volume. Authority Score figure from a separate Semrush study of 1,000 domains. Correlation, not causation.
The headline finding, and what it is not
Across every independent dataset published in the last year, unlinked brand mentions outrank link volume as a correlate of AI visibility. Brands in the top quartile by mention volume receive roughly 10x more AI citations than the rest of the field.
What this does not establish is causation. High AI visibility co-occurs with a wide cross-platform presence. Buying mentions is not a demonstrated mechanism for producing citations, and anyone selling it as one is overreading the data. The defensible read is that being genuinely discussed across the web is a prerequisite for being retrieved, and mention volume is the cheapest available measure of that.
Third-party coverage, not owned properties, supplies most of what AI systems cite. Journalism alone accounts for 20% to 30%.
The same story syndicated across multiple publications generated up to 325% more citations than publishing it only on the owned site.
Content updated within 30 days. Pages untouched for three months are 3x more likely to lose citations they already had.
44.2% from the opening 30%, 31.1% from the middle, 24.7% from the final third. Front-load the answer.
Three findings worth dwelling on
YouTube is a retrieval surface, not a marketing channel
YouTube mentions show the strongest single correlation in the dataset at 0.737, and YouTube supplies 23.3% of all AI Overview citations, roughly a 200-fold advantage over competing video sources. Some of that is a Google-owned-property artifact and should be discounted accordingly. Enough of it survives the discount that publishing video versions of your highest-value written assets is a defensible AI visibility tactic rather than a brand exercise.
Content position beats content volume
44.2% of extracted citations come from the opening 30% of a document. Pages over 20,000 characters show a 4.3x citation advantage over thin content, but length is not the mechanism. Depth produces more retrievable passages; padding produces more words. A 20,000-character page that buries its conclusions performs worse than a 4,000-character page that answers cleanly in the first screen.
Structure is cheap and measurable
68.7% of cited pages maintain a clean H1 to H2 to H3 hierarchy without skips. Organization schema appears on 25% of ChatGPT-cited pages and 34% of AI Mode-cited pages; Article schema on 20% and 26%; BreadcrumbList on 15% and 20%. None of these are silver bullets. All of them are an afternoon of work.
A note on backlinks in the AI era
Backlinks have not stopped mattering, their role changed shape. Analysis of 1,000 domains found the relationship between authority and AI mentions far stronger by rank correlation than by linear correlation, which is the signature of a threshold rather than a slope: clear an authority floor to be considered, after which additional links buy comparatively little. The same study found follow and nofollow links correlate almost identically, and image links slightly outperform text links. AI systems appear to read links as evidence that a brand is discussed, not as votes that accumulate.
The GEO playbook
Generative engine optimization techniques ranked by the strength of evidence behind them, drawn from controlled benchmarks rather than agency case studies. Some of this reproduces reliably. A surprising amount of the popular advice does not.
GEO technique effectiveness in controlled benchmarks
Change in position-adjusted citation share. Bars below the line tested neutral to harmful.
Scale -25% to +50%. Tag shows confidence in the finding.
Source. Synthesized from the 2026 critical survey of generative engine optimization, the original KDD 2024 GEO paper, C-SEO Bench and SAGEO Arena end-to-end testing. Effects are conditional on the pipeline stage noted in the reference table below.
What holds up
Answer first, evidence second
High confidence
44.2% of citations are extracted from the first 30% of a page, and models strongly favor passages that explicitly align with the question asked.
- Open each section with a direct, complete answer in two or three sentences
- Write so a single paragraph makes sense lifted out of the page entirely
- Put caveats and methodology below the answer, never before it
- Use the question as the H2, phrased the way a person would actually ask it
Build the off-site mention footprint
High correlation causality unproven
Mentions correlate at 0.66 to 0.71 against 0.22 for raw backlinks. Earned media supplies roughly 82% of AI citations.
- Pursue coverage and commentary placements, not only link placements
- Treat an unlinked mention in a credible publication as a win worth paying for
- Syndicate substantive content rather than hoarding it on the owned site
- Keep entity details identical everywhere so one brand resolves to one entity
Supply verifiable, attributed evidence
Moderate, +20% to +27%
Statistics, definitions, direct quotations and named sources consistently raise the share of an answer a document wins, provided they are real.
- Attach a source and a date to every number you publish
- Prefer first-party data no competitor can replicate
- Quote named experts rather than paraphrasing anonymously
- Never invent a statistic to be quotable. It is a short gain and a lasting cost
Treat freshness as maintenance
High, 3.2x multiplier
Pages updated within 30 days earn 3.2x the citations. Pages untouched for three months are 3x more likely to lose citations they held.
- Run a quarterly refresh on your highest-value pages before writing anything new
- Update the substance, not the date stamp. Cosmetic changes do not earn the multiplier
- Expose an accurate dateModified in structured data
- Retire or consolidate pages you are not willing to maintain
Write for fan-out, not the head term
Moderate, mechanistically grounded
AI Mode decomposes one prompt into many sub-queries. Prompts average 7.22 words against 4.0 for classic search.
- Map the sub-questions a prompt fans out into, then cover each in its own section
- Build comparison content. Comparison articles take 32.5% of citations in tracked datasets
- Cover the awkward adjacent questions competitors skip: pricing, limits, drawbacks
- Write in the register people use with an assistant, not with a search box
What to stop doing
These tested neutral or negative
Only 3 of 54 method-and-domain combinations in one benchmark showed a significant positive effect.
- Keyword stuffing for LLMs. Reduces position-adjusted visibility. Does not transfer from classic SEO
- llms.txt as a visibility play. 97% of published files are never requested at all
- Hidden instructions aimed at the model. Prompt injection in page copy is manipulation, and detectable
- Generic formatting rituals. Tables and bullets everywhere generalizes poorly across domains
- Body-only rewrites at the expense of relevance. Measurably reduced retrieval end to end
Where optimization becomes manipulation
The research literature converges on four cumulative tests. A technique passing all four is optimization. A technique failing any one of them is manipulation, and carries the same class of risk as every other tactic that works until it is detected.
Semantic preservation
The claims remain true after the rewrite. If optimizing the page changed what the page asserts about the world, you did not optimize it.
Evidentiary authenticity
Every statistic, quotation and reference is real and verifiable. Adding a fabricated number may increase reuse in the short term while degrading the thing that makes citation valuable in the first place.
Content and instruction separation
The page contains no text written to command the model rather than inform the reader. Hidden prompts, invisible text and instruction-shaped copy all fail here regardless of whether they currently work.
Disclosure
Commercial intent is visible and competitors are represented fairly. A comparison page that exists to win a comparison query should still be a comparison a reader could rely on.
The uncomfortable finding nobody sells
Controlled multi-actor experiments show individual GEO gains eroding toward zero as adoption spreads. The benchmark language for this is that it approaches a zero-sum game: early adopters capture a redistribution of existing visibility rather than creating new visibility. Combined with the reliability problem in chapter 09, that argues for investing in the durable inputs (genuine expertise, first-party data, earned coverage) over the tactical ones, because the tactical edge has a half-life and the durable one does not.
Platform fragmentation
Google, ChatGPT, Gemini and Perplexity answer the same question from largely different sources. There is no global AI ranking to optimize toward, and a win on one surface predicts very little about the others.
URL-level overlap between Google organic, AI Overviews and Gemini runs 0.11 to 0.18. Near-disjoint source sets.
Bing Chat and Perplexity share about 26%. Every pairing is lower than practitioners assume.
Across 680 million analyzed citations. Extreme concentration at the top of the field.
Globally, every month, in the top 100. YouTube, Google, Reddit, Amazon, Apple, Walmart, Disney and a handful of others.
What the fragmentation means in practice
Concentration cuts both ways
The top 15 domains capture roughly 68% of all AI citations. YouTube alone supplies 23.3% of AI Overview citations, Wikipedia 18.4% and Reddit around 21%. For most brands the realistic route into an answer runs through those platforms rather than around them. Being discussed on Reddit and present on YouTube is a distribution strategy, not a vanity exercise.
Vertical dynamics vary enormously
In news and media the top three brands hold 82.9% of AI visibility. In consumer electronics it is 76.9%. Finance sits at 41.4% and industrial at 42.2%. If you operate in a concentrated vertical, displacing an incumbent is close to impossible in the short term and the winnable game is long-tail and comparison prompts. In a distributed vertical there is genuine room to move, and the effort is better spent.
Mention and citation are different wins
On Gemini the overlap between brands mentioned in an answer and domains cited beneath it can be as low as 30%. Brand visibility across repeated AI sessions is itself only about 30% consistent when the same prompt is run again. Being named in the answer and being the source of the answer are separate outcomes that need separate tracking, and most tooling conflates them.
The practical consequence
Pick the one or two surfaces where your audience actually is and measure those properly, rather than buying a dashboard that averages five platforms into a single visibility score. An average across surfaces with 11% source overlap is not a measurement of anything. If your buyers research in ChatGPT, a Gemini score is decoration.
Crawlers and access
What the bots take, what they return, and how to make an access decision with numbers instead of instinct. Bots now account for 57.5% of HTML traffic, which makes this a bandwidth question as well as a visibility one.
Crawl-to-referral ratio by bot
Pages crawled per single visit returned. Logarithmic scale.
Source. Cloudflare Radar, Q1 to Q2 2026. ClaudeBot improved from 23,951:1 in Q1 to 11,122:1 by late spring, so read these as a moving target rather than a fixed rate.
The access decision, with numbers
AI crawlers are roughly 20.3% of verified bot traffic with another 6.5% from AI search bots. Blocking is a real lever and an increasingly common one, but it is not a single decision. Training crawlers and answering crawlers do different jobs and deserve different answers.
The ratios are lopsided but improving
ClaudeBot crawled roughly 23,951 pages per referred visit in Q1 2026, improving to about 11,122 to 1 by late spring. GPTBot sat near 1,276 to 1, PerplexityBot 111 to 1, Microsoft Copilot 33 to 1. Google runs about 4.9 to 1 and DuckDuckGo about 1.5 to 1. The trend is moving the right way. The asymmetry is still two to three orders of magnitude.
Separate training from answering
As of May 2026, 51.8% of AI crawler activity was training-focused, 35.7% mixed and 9.3% search-only, up from 7.5% a month earlier. The emerging publisher consensus is to disallow training crawlers while allowing the answering and retrieval bots that can actually cite you. Blocking indiscriminately removes you from answers without recovering the bandwidth that matters.
Blocking went mainstream
More than 2.5 million sites fully disallow AI training crawlers, and over a million activated network-level blocking after default protections shipped in mid-2025, which blocked roughly 416 billion scraping requests in five months. In robots.txt files, GPTBot appears in 5.52% of disallow rules, CCBot in 5.08% and ClaudeBot in 4.88%.
llms.txt is not the answer
A study of 137,210 domains found 28% publish an llms.txt file and 97% of those received zero requests. Of the small remainder with any traffic, 96% came from bots, 77% of those from non-AI tooling such as SEO auditors, and AI retrieval bots accounted for just 1.1% of requests. Publish it if your platform generates one. Do not build a workflow around it.
Technical checklist for AI retrieval
- Server-render your primary content. Several AI crawlers execute little or no JavaScript. Content that only exists after hydration may never be retrieved at all, and this is now one of the most common silent failures.
- Audit robots.txt deliberately. Decide bot by bot and document the decision. A large number of sites are blocking answering crawlers by accident through a broad wildcard rule written years ago for a different problem.
- Ship Organization, Article and BreadcrumbList schema. Present on 25% to 34% of cited pages, higher in AI Mode than ChatGPT. Cheap, and it disambiguates your entity.
- Keep dateModified honest. Freshness carries a 3.2x citation multiplier, and a false timestamp on unchanged content is a spam signal rather than a shortcut.
- Make every section independently retrievable. Self-contained headings, no pronoun chains reaching back to earlier sections, no critical context trapped inside an image.
- Do not rate-limit or CAPTCHA the answering bots you want. Verify the ones you allow rather than blocking broadly and hoping the right ones get through.
- Watch server logs, not just analytics. AI crawl activity is invisible in client-side analytics. Server logs are the only honest record of who is actually reading you.
Measuring AI visibility
AI answers are stochastic. A large share of the reporting built on top of them is measuring noise and calling it progress. This chapter is the part of the report most likely to save you money.
The reliability problem
How much citation selection changes within 24 hours on the same prompt, even at temperature zero.
Minimum repetitions per prompt before a measurement means anything, per current research practice.
Consistency of brand visibility across repeated AI sessions with identical input.
Of generated sentences fully supported by the citation attached to them. Visibility is not accuracy.
A screenshot is not a measurement. If your agency, your tool or your own reporting checks a prompt once and reports the result, the finding is indistinguishable from chance at the effect sizes anyone is claiming. Run each prompt 7 to 8 times across multiple paraphrases, and report a rate with a range rather than a binary.
Measure the right layer
A single "AI visibility score" hides more than it reveals. These six layers fail independently, and fixing the wrong one wastes a quarter.
| Layer | Question it answers | How to measure it | What it cannot tell you |
|---|---|---|---|
| Activation | Does this query trigger an AI answer at all? | SERP feature tracking across your keyword set | Anything about your own visibility |
| Retrieval | Did your page make the candidate pool? | Server log analysis of AI crawler hits by URL | Whether you will be selected or cited |
| Citation | Were you named as a source? | Repeated prompt sampling, 7 to 8 runs, multiple paraphrases | Whether the claim attributed to you was accurate |
| Prominence | How much of the answer came from you? | Position-adjusted share of the generated response | Whether a human actually read that part |
| Sentiment | How is your brand characterized? | Classified sampling of mentions across prompt sets | Commercial outcome |
| Behavior | Did any of it produce revenue? | Referrer-segmented analytics plus self-reported attribution | Causality, without a controlled test |
A protocol you can defend in a meeting
- Define the prompt set before you measure. 40 to 60 prompts covering the questions your buyers actually ask, written in assistant register rather than keyword register. Freeze it, so quarter-over-quarter comparisons mean something.
- Paraphrase each prompt three ways. Fan-out means small wording changes route to different sub-queries. A single phrasing measures one path through the system.
- Run each variant at least seven times. Record a citation rate, not a yes or no. Report the range alongside the mean.
- Track activation separately. If 30% of your prompt set does not trigger an AI answer at all, your citation rate has a different denominator than you think it does.
- Pull server logs monthly. Retrieval failure and citation failure look identical in a dashboard and need completely different fixes. Logs are the only place they separate.
- Hold one cohort untouched. If you change nothing on 20% of your pages, you have a control group, which is the only thing that turns a correlation into evidence about your own site.
Analyst note
Only about 23% of teams investing in generative engine optimization are measuring it. That gap is the actual opportunity in this report. Most of your competitors are running GEO programs they cannot evaluate, which means they cannot tell a real gain from drift and will keep optimizing toward whatever their last screenshot happened to show. A mediocre strategy with honest measurement beats a sophisticated one without it, because only one of the two can correct itself.
Industry and query variation
Aggregate weights are a starting point, not a plan. Emphasis shifts sharply by vertical and by query type, and the AI layer has widened those differences rather than smoothing them.
Three patterns worth acting on
YMYL still behaves differently, and now behaves differently twice
Healthcare and finance have always carried heavier E-E-A-T weighting. In 2026 they also sit at opposite ends of the AI citation spectrum: healthcare retains the highest remaining organic overlap at 24% while finance has the lowest at 11%. A finance page that ranks well is telling you almost nothing about whether it will be cited. A healthcare page that ranks well is telling you a little.
E-commerce got a reprieve, not a reversal
AI Overview coverage in e-commerce fell from 29% to roughly 4% as Google routed commercial intent toward Shopping surfaces and ad units. If you sell products, generative answers are currently a minor factor in your acquisition mix. That is a current condition rather than a structural one, and agentic purchasing is the obvious place it changes.
Informational content is where the citation upside lives
Comparison articles take 32.5% of tracked citations and opinion pieces about 10%. Informational and consideration-stage queries are where fan-out concentrates, which is exactly the content most commercial sites under-invest in because it does not convert on the session. In an AI-mediated funnel it converts later, through a different door.
Device still matters, just less interestingly
Mobile accounts for roughly 60% to 65% of search traffic, and 77.2% of mobile searches now end without a click against 60% overall. Mobile-first indexing made mobile-friendliness table stakes rather than an advantage, so the useful mobile question in 2026 is no longer whether your pages work on a phone. It is whether the zero-click rate on your mobile queries has quietly made a channel you still report on commercially irrelevant.
Priorities and quick wins
Everything above is diagnosis. This is the part you can act on next week, ordered by expected return against effort rather than by how interesting it is to work on.
Impact against effort
Q1 Do these first
High impact, low effort
- Audit robots.txt for accidental AI crawler blocks. Same day, and a meaningful number of sites are silently excluded from answers by a wildcard rule
- Front-load a direct answer in every section. Two to three sentences before the context. 44% of citations come from the top third
- Refresh your top 20 pages before writing anything new. 3.2x citation multiplier within 30 days
- Ship Organization, Article and Breadcrumb schema. One afternoon, present on a third of cited pages
- Segment reporting by AI Overview presence. Stops you diagnosing a visibility problem as a quality problem
Q2 Plan and resource these
High impact, high effort
- Build an earned mention program. The single strongest correlate of AI visibility, and the slowest to build
- Publish first-party data nobody else has. The one input models cannot synthesize from their own priors
- Server-render primary content. Expensive on some stacks, and the difference between retrievable and invisible
- Stand up a real measurement protocol. 40 to 60 frozen prompts, three paraphrases, seven runs. See chapter 09
- Publish video versions of top written assets. YouTube correlates at 0.74 and supplies 23.3% of AI Overview citations
Q3 Fill-in work
Lower impact, low effort
- Tighten heading hierarchy to a clean H1 to H2 to H3 with no skips
- Shorten URL slugs toward the 17 to 40 character range
- Add internal links from strong pages to the ones you want retrieved
- Keep dateModified accurate rather than aspirational
- Add alt text and captions that carry the actual information in a figure
Q4 Do not bother
Low or negative return
- llms.txt. 97% of published files are never requested
- Keyword stuffing for models. Reduces position-adjusted visibility
- Hidden instructions to the model. Manipulation, detectable, and a liability with a countdown on it
- Generic formatting rituals. Tables and bullets everywhere generalizes poorly
- Averaging five platforms into one visibility score. Source overlap is 11%. The average measures nothing
- Chasing citation on a surface your buyers do not use
Where the effort should go
If you are reallocating a budget rather than adding to one, this is the shape the 2026 data supports. Content and authority still dominate, but the authority slice increasingly means mentions rather than links, and a measurement line is now non-optional.
The organizational finding
Teams running SEO and AI visibility as one integrated workflow reported increased traffic or leads 81% of the time. Teams running them as separate initiatives reported the same result 36% of the time. The split itself appears to be the thing that costs you, which makes "who owns AI visibility" a more consequential question than most of the tactics in this chapter.
What still goes wrong
The classic failures have not gone away. They have been joined by a set of new ones that are specific to an AI-mediated search layer, and the new ones are harder to see because nothing in a standard dashboard reports them.
New in 2026
Diagnosing a visibility drop as a quality problem
Impressions up, positions stable, clicks down. That is an AI Overview absorbing the click, not a content failure. Teams rewrite healthy pages, see no recovery, and conclude the algorithm is broken.
Fix: segment by AI Overview presence and compare cohorts against each other.
Client-side rendering your primary content
Several AI crawlers execute little or no JavaScript. A page that renders perfectly for users and Googlebot can be entirely absent from the candidate pool for an AI answer.
Fix: server-render the substance. Check what a plain fetch returns.
Blocking answering bots with a wildcard
A broad robots.txt rule written years ago for scrapers now excludes the retrieval bots that would have cited you, while doing nothing about the training crawlers you actually object to.
Fix: decide bot by bot. Disallow training, allow answering.
Measuring AI visibility from screenshots
Citation selection drifts 9% to 28% within a day. A single check reports noise. Teams celebrate and panic in alternating weeks over movements that are not real.
Fix: 7 to 8 runs per prompt, multiple paraphrases, report a rate.
Publishing new content while the best pages go stale
Pages untouched for three months are 3x more likely to lose citations they already held. Most editorial calendars still reward publication volume over maintenance.
Fix: quarterly refresh cycle on top pages, booked before new commissions.
Buying an averaged multi-platform visibility score
Source overlap between ChatGPT and Perplexity is about 11%. Averaging five surfaces into one number produces something that cannot go up or down for any diagnosable reason.
Fix: measure the one or two surfaces your buyers actually use.
Still true from previous editions
- Thin or duplicated content at scale. Generative tooling made this cheaper to produce, which made the penalty for it more consistently applied. Five spam updates across 2025 and 2026 is not a coincidence.
- Bulk link acquisition. Now doubly wasted: it was already a weak ranking signal at volume, and it correlates with AI visibility at 0.22. The money buys a number in a tool.
- Keyword stuffing. Did not work for rankings, actively reduces position-adjusted visibility in AI answers. It is one of the few tactics that is clearly negative in both systems.
- Neglecting Core Web Vitals until they fail. A threshold factor. Nobody wins by being fast, plenty lose by being slow.
- Publishing expertise you cannot demonstrate. Author identity, credentials and first-hand testing are what separate real expertise from fluent text, and that distinction got more valuable the moment fluent text became free.
- Treating a traffic number as the goal. Sessions are down across most of the web while AI referrals convert at several times the organic rate. A channel report that only counts visits will tell you to abandon the most valuable traffic you have.
How we got here
Ranking signals do not lurch. They drift, and the drift only looks like a lurch in hindsight. Eight years of movement, then the confirmed update log for the current window.
Signal influence over time, 2018 to 2026
Relative influence score. AI citation visibility did not exist as a measurable surface before 2023.
View the underlying numbers
| Signal | 2018 | 2019 | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 | 2026 |
|---|---|---|---|---|---|---|---|---|---|
| Content quality | 65 | 70 | 73 | 76 | 79 | 82 | 84 | 86 | 87 |
| Authority signals | 68 | 70 | 71 | 72 | 73 | 75 | 79 | 83 | 85 |
| Brand and entity | 40 | 42 | 44 | 47 | 51 | 56 | 64 | 74 | 82 |
| User experience | 35 | 40 | 45 | 50 | 55 | 59 | 62 | 65 | 66 |
| AI citation visibility | n/a | n/a | n/a | n/a | n/a | 8 | 24 | 56 | 79 |
| Technical SEO | 45 | 44 | 43 | 42 | 41 | 40 | 39 | 38 | 41 |
Source. Composite index built from published correlation studies across the period. Directional trend, not a measured absolute.
The citation era. Ranking and visibility split apart.
AI Overviews reached roughly 48% of US queries, up from 6.5% in January 2025, and AI Mode established itself as a distinct surface where 92% to 94% of sessions end without an outbound click. The defining change was structural rather than algorithmic: the share of AI Overview citations drawn from the organic top 10 collapsed from around 76% in mid-2024 to between 17% and 38%. Off-site brand mentions overtook raw backlink volume as the strongest published correlate of AI visibility.
AI content era and the AI Overviews rollout
Generative answers moved from experiment to default surface. Helpful content principles were folded into the core system rather than run as a separate classifier. Authority signals surged in prominence from 2023 onward, with top-performing pages consistently outperforming purely UX-optimized sites in commercial results. Google's position stayed consistent: helpful content is rewarded regardless of how it was produced, and originality, accuracy and first-hand experience are what separate the two.
Page experience and Core Web Vitals
Core Web Vitals arrived as lightweight ranking signals combining loading, interactivity and visual stability. Mobile-first indexing completed globally. The Helpful Content Update began prioritizing content made for people over content made for search engines, which in retrospect was the beginning of the shift the AI layer completed.
E-E-A-T and the quality turn
Quality rater guidelines formalized experience, expertise, authoritativeness and trust, with the heaviest application to Your Money or Your Life topics. BERT brought genuine natural language understanding to query interpretation, which started the long decline of exact-match keyword weighting that is still continuing in 2026.
Forecast to 2028
Forecasts in this field age badly, so these are framed as bets with stated reasoning rather than predictions with implied certainty. Each one names what would prove it wrong.
Six bets for 2027 and 2028
Brand and entity authority becomes the dominant input
Confidence: high. Every independent dataset points the same way, and the mechanism is coherent: models resolve questions to entities before they resolve them to documents. Mention volume is currently the cheapest proxy for entity recognition anyone has.
Falsified if: a major surface ships retrieval that weights document-level relevance over entity familiarity, and mention correlation drops below link correlation in a replicated study.
Citation readiness matters more than ranking for informational content
Confidence: high. At 48% AI Overview coverage and 92% AI Mode zero-click, the citation is the impression for a growing share of informational queries. The organic position is increasingly a means to retrieval rather than an end.
Falsified if: AI Overview coverage plateaus or reverses under regulatory pressure, or click-through on generative surfaces recovers toward traditional SERP levels.
Passage-level structure becomes a formal discipline
Confidence: moderate to high. Retrieval operates on chunks. As soon as measurement gets good enough to attribute citation to specific passages, content teams will start optimizing the chunk rather than the page, and briefs will change shape.
Falsified if: long-context retrieval makes chunking irrelevant and whole-document relevance dominates selection.
First-party data becomes the primary durable moat
Confidence: moderate to high. When fluent text is free, the scarce input is information that exists nowhere else. Models cite what they cannot synthesize from their own priors, and proprietary benchmarks and survey data are exactly that.
Falsified if: synthesis quality improves enough that models reconstruct proprietary findings from adjacent public sources without citing the original.
Multimodal presence stops being optional
Confidence: moderate. YouTube already correlates at 0.74 and supplies 23.3% of AI Overview citations. Some of that is a Google-owned-property artifact, which is why the confidence here is lower than the raw number suggests.
Falsified if: the YouTube advantage narrows sharply once discounted for platform ownership, or non-Google surfaces continue to weight video lightly.
Agentic and transactional readiness becomes the next contested surface
Confidence: moderate, timing uncertain. Assistants are moving from answering to acting. Google still holds roughly 90% of transactional queries, and e-commerce AI Overview coverage actually fell to about 4%, so this is a forecast about 2027 to 2028 rather than a current condition.
Falsified if: agent-mediated purchasing stalls on trust, payments or liability, which is the most likely outcome in regulated categories.
The adoption gap is the opportunity
Plan to optimize for AI search
Planning to start within 3 to 6 months
Actively implementing today
Actually measuring the result
Nearly everyone intends to do this. Under a quarter can tell whether it worked. Given that repeated prompts change citation selection by 9% to 28% within a day, a team without a sampling methodology cannot distinguish a real gain from drift, and will keep optimizing toward whatever their last check happened to show.
Forecast note from Jourdan Rombrough
The next two to three years will be decided by whether a brand is discussed, not just linked. Every reproducible finding in the AI visibility research points the same direction: models retrieve entities they recognize, then cite passages that answer the question cleanly and carry verifiable evidence. That rewards most of what good SEO already rewarded, with the weighting rearranged and the feedback loop made considerably noisier.
My advice for 2027 planning is unglamorous. Fix measurement first, because without it every other decision is a guess. Invest in earned mentions and first-party data second, because those are the inputs with no half-life. And treat any tactic that works only because a model can be tricked as a liability with a countdown on it, because five spam updates in eighteen months is a fairly clear statement of intent.
Questions people actually ask
The questions that come up in every strategy conversation this year, answered with the data rather than with reassurance.
Do I still need traditional SEO if AI search is taking over?
Yes, and the framing of the question is the trap. Google still handles roughly 57% of global digital queries against ChatGPT's 17.9%, and its lead in transactional queries is close to 90%. More importantly, the signals that earn AI citations overlap heavily with the signals that earn rankings: content quality, topical depth, freshness, demonstrable expertise and clean structure appear on both lists.
What changed is the weighting and the reporting. Off-site brand mentions now correlate with AI visibility roughly three times more strongly than raw backlink volume, and between 62% and 83% of AI Overview citations come from pages outside the organic top 10. Teams running SEO and AI visibility as one integrated program reported increased traffic or leads 81% of the time, against 36% for teams running them separately. Run one program, not two.
What is generative engine optimization, and does it actually work?
GEO is optimizing to be retrieved and cited by AI answer systems rather than ranked in a list of links. Parts of it are well evidenced and parts of it are marketing. What reproduces across independent studies: explicit query-intent alignment, verifiable attributed evidence such as real statistics and quotations (20% to 27% gains in benchmarks), content recency (a 3.2x citation multiplier within 30 days), clean passage-level structure, and an off-site mention footprint.
What fails to reproduce: keyword stuffing aimed at models (null to negative), generic formatting rituals (only 3 of 54 method-and-domain combinations tested significantly positive in one benchmark), and llms.txt as a visibility play. Be skeptical of any large headline number. The widely quoted 40% gain from the original GEO research was measured on documents that had already been retrieved, and end-to-end testing including retrieval found body-only rewrites can reduce top-10 presence by around 16%.
My rankings are flat but traffic is down. What happened?
Check impressions against clicks before you touch a single page. Across publishers, total impressions rose roughly 49% since AI Overviews launched while click-throughs fell about 30%. An AI answer now sits above your result and absorbs the click.
Click rate on queries with an AI Overview measured 8% against 15% without, and only about 1% of users click a link inside the overview itself. Roughly 60% of US searches now end with no click at all, rising to 77.2% on mobile. Segment your keyword set by whether an AI Overview is present and compare the two cohorts. Then check whether you are being cited, because cited brands earn roughly 120% more organic clicks per impression than uncited ones on the same queries.
Are backlinks still important?
Yes, but as a threshold rather than a scoreboard. Analysis of 1,000 domains found the relationship between authority and AI mentions far stronger by rank correlation than by linear correlation, which is the statistical signature of a floor you have to clear rather than a slope you climb. Once you are over it, additional links buy comparatively little.
Two findings from the same study are worth sitting with: follow and nofollow links correlate almost identically with AI visibility, and image links slightly outperform text links. Read together, that suggests AI systems treat links as evidence a brand is discussed rather than as votes that accumulate. Which is also why unlinked mentions correlate at 0.66 to 0.71 while raw backlink count sits at 0.22.
Should I block AI crawlers from my site?
It is not one decision, it is two. Training crawlers and answering crawlers do different jobs, and blocking indiscriminately removes you from AI answers without recovering much. As of mid-2026 roughly 51.8% of AI crawler activity was training-focused, 35.7% mixed and 9.3% search-only.
The economics are genuinely lopsided: ClaudeBot crawled on the order of 11,000 pages per referred visit, GPTBot around 1,276 to 1 and PerplexityBot 111 to 1, against roughly 4.9 to 1 for Google. But the traffic that does arrive converts well, at 15.9% for ChatGPT and 10.5% for Perplexity against a Google organic average near 1.76%. The emerging publisher consensus is to disallow training crawlers while allowing retrieval and answering bots, decided bot by bot rather than with a wildcard.
Do I need an llms.txt file?
On current evidence, no. A study of 137,210 domains found that 28% publish an llms.txt file and 97% of those received zero requests. Among the small remainder that saw any traffic, 96% came from bots, 77% of those from non-AI tooling such as SEO auditors and tech profilers, and AI retrieval bots accounted for just 1.1% of requests.
If your CMS generates one automatically, leave it. Do not build a workflow around it, do not bill a client for it, and do not report it as an AI visibility initiative. The time is better spent server-rendering your content, shipping Organization and Article schema, and auditing which crawlers your robots.txt is accidentally blocking.
How do I measure AI visibility without fooling myself?
Start by accepting that AI answers are stochastic. Repeated runs of the same prompt change citation selection by 9% to 28% within 24 hours even at temperature zero, and brand visibility across repeated sessions is only about 30% consistent. Research practice is a minimum of 7 to 8 repetitions per prompt across multiple paraphrases before any number means anything.
Then measure the right layer, because they fail independently: activation (does the query trigger an AI answer at all), retrieval (are AI crawlers hitting the URL, visible only in server logs), citation (are you named as a source), prominence (how much of the answer came from you), sentiment (how you are characterized) and behavior (referrer-segmented revenue). Only about 23% of teams investing in GEO are measuring it at all, which is precisely where the advantage is.
How much does content freshness actually matter now?
More than any other single content lever. Pages updated within 30 days earn 3.2x the AI citations of older equivalents, and pages untouched for three months or more are 3x more likely to lose citations they already held. Roughly 65% of AI Overview citations come from content under a year old and 89% from content under three years old. On ChatGPT, 76.4% of the most-cited pages had been updated in the previous 30 days.
The practical implication is a change in how a content team spends its week. Refreshing your twenty best pages now has a higher expected return than publishing three new ones. Update the substance though, not the date stamp. A false timestamp on unchanged content is a spam signal, not a shortcut.
Is AI-generated content penalized?
Not for being AI-generated. Google's stated position is consistent and has not changed: helpful content is rewarded regardless of how it was produced, and what matters is originality, accuracy, first-hand experience and demonstrated expertise. What has changed is the enforcement rate around the failure mode that generative tooling makes cheap. Five spam updates across 2025 and 2026, against four core updates, is a fairly direct statement about scaled low-effort content.
The AI-era version of the advice is sharper than the old one. When fluent text is free, fluency stops being a differentiator and the scarce inputs become the things a model cannot generate: proprietary data, first-hand testing, named humans with verifiable credentials, and findings that are inconvenient enough that nobody would have invented them.
How long should I wait before judging any of this?
For ranking changes, the honest answer has not changed: three to six months for content and authority work on a site with reasonable existing equity, longer on a new domain, and core updates arrive on their own schedule regardless of what you did. There were four core updates between March 2025 and June 2026, so a twelve-month window will usually contain at least two events you did not control.
For AI citation, the measurement noise dominates at short horizons. Given 9% to 28% drift within a single day, a month is the minimum useful comparison window and a quarter is better. Hold one cohort of pages untouched so you have a control, because without it you will attribute a platform-side change to your own work in whichever direction flatters the last thing you did.
Methodology and limits
What this report is, what it is not, and every source behind it. If a number here cannot be traced to a study with a stated sample, it should not be in the report and you should tell me about it.
What was done
Collection window. October 2025 to September 2026, with correlation figures drawn from studies covering the January 2025 to January 2026 period and sampling more than a million URLs.
Method. Aggregation of published correlation studies, AI citation datasets, crawler telemetry, behavioral research and the 2026 academic survey literature on generative engine optimization, cross-checked against Google's own documentation and public statements. Weight percentages are estimated relative influence normalized to 100, not measured values. Google publishes none of this.
Where studies disagree, the disagreement is shown with both methodologies rather than averaged into a single number, because the reason for the divergence is usually more informative than either figure. AI Overview prevalence is the clearest example: reported coverage ranges from 16% to 50% almost entirely as a function of keyword set composition.
What this report is not
- Not a leak or an insider account. Nothing here comes from Google's ranking systems directly. It is inference from observed behavior.
- Not causal. Nearly every figure is correlational. High AI visibility co-occurs with a wide mention footprint; that does not establish that buying mentions produces citations.
- Not globally representative. The underlying studies skew heavily toward English-language, Western-market, commercially valuable queries, because that is where the measurement tooling points. Treat findings as least reliable where your market is least like that.
- Not stable. Four core updates and five spam updates landed inside this report's window. Any coefficient here is a snapshot, not a constant.
- Not vendor-neutral in its inputs. Many of the largest AI citation datasets are published by companies selling AI visibility tooling. They are directionally useful and worth discounting on magnitude, and they are labeled throughout.
Sources
- Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization, 2023 to 2026.
- GEO: Generative Engine Optimization.
- C-SEO Bench and SAGEO Arena.
- Ahrefs, AI visibility correlation study.
- Ahrefs, llms.txt analysis.
- Semrush, technical SEO and AI search study.
- Semrush, 2026 AI Visibility Index.
- Semrush, backlinks and AI search visibility.
- Semrush, Google AI Mode early adoption and SEO impact.
- BrightEdge, AI Overview coverage tracking.
- Seer Interactive, AI Overview CTR longitudinal study.
- Pew Research Center, search behavior study.
- SparkToro and Datos, zero-click research.
- Similarweb, publisher traffic and AI referral tracking.
- Cloudflare Radar, bot and AI crawler telemetry.
- Surfer SEO, AI Overview citation analysis.
- Muck Rack, AI citation and earned media study.
- ConvertMate, citation freshness analysis.
- Adobe Digital Insights, AI referral commerce data.
- Zyppy, 2026 Google ranking factors expert survey.
- Google Search Central and the Search Status Dashboard.
- Search Engine Land and Search Engine Journal.
Found an error?
If a figure here is wrong, out of date, or has been superseded by better research, I would rather know. Corrections go to optimizationtheory.com and get reflected in the next build with the change noted. This report is versioned and rebuilt rather than quietly edited.