$ cat CTR & NavBoost Manipulation
NavBoost and click evidence from public testimony and leak analysis; why click-use evidence does not establish purchased-click efficacy.
CTR manipulation attempts to manufacture search clicks or engagement to influence rankings. Public antitrust testimony in 2023 already described NavBoost’s use of click information; the 2024 documentation leak supplied a different, less complete view of fields and terminology. Neither establishes that purchased clicks reliably cause durable ranking gains. This chapter separates system evidence from vendor efficacy claims. [3]
1 · Definition
CTR manipulation is the artificial generation of search clicks and post-click engagement, real or simulated, to make a page look more attractive and satisfying to searchers than it is, in hopes that Google's click-based re-ranking promotes it. "NavBoost" is the internal Google system, named in the leak, that records click signals. leak-indicated[1]
2 · Why it might work (carefully)
Analysts describe leaked click-related fields such as goodClicks, badClicks and lastLongestClicks. Fishkin and King offer interpretations, not an independently audited production specification. In particular, readings of “squashing” differ: normalization of a strong signal and anti-spam filtering are not the same mechanism. The archive cannot resolve which operations apply to a current query or identify a purchase volume that changes rankings. [1][2]
Public court testimony supports that NavBoost exists and uses click information. It does not establish that a third party can manufacture useful signals or produce lasting rank improvements. No controlled purchased-click results are supplied here. [3]
3 · Architecture
CLICK SOURCE ─▶ SEARCH FOR TARGET QUERY ─▶ CLICK TARGET RESULT ─▶ DWELL / INTERACT ─▶ (hope) NAVBOOST +
· bot farm (residential proxies, (skip competitors, (scroll, second
· crowd (paid varied geos/devices) "last longest click") page, no bounce)
humans) │
└──────────── Proposed causal path only; filtering and production behaviour are not established here ──┘
4 · Historical evolution
| Era | What changed |
|---|---|
| 2014–2019 | Rand Fishkin's "IRL CTR" experiments spark debate; Google publicly downplays clicks. Folklore era. |
| 2020–2023 | CTR bots and crowd-click services are discussed in practitioner accounts; this archive has no time-series census to establish proliferation. disputed |
| 2023 | Public antitrust testimony discusses NavBoost and click information. [3] |
| 2024 | Analysts report NavBoost-related fields in leaked documentation. Field interpretation is separate from the earlier testimony. [1] |
| 2024–2026 | Renewed discussion; this archive supplies no reproducible estimate of purchased-click effectiveness. |
5 · Infrastructure and evidence boundaries
Services may advertise paid crowds or browser automation. Those are advertised delivery models, not a representative survey of typical operators. The defensive workflow below uses authorized exports and a deployment log. No device farm, proxy rotation or manufactured engagement is needed to diagnose a measurement claim, and none has a measured ranking return in this archive.
6 · Setup & prerequisites
For an observational study, preserve Search Console exports by page, query, device, country and date. Record impressions as well as clicks: CTR is clicks divided by impressions, while average position aggregates changing result contexts. A different query mix can change both metrics without any effect from a title edit.
7 · Workflow overview
Record a stable baseline → select comparable untreated pages → change a documented title/snippet treatment → collect repeated exports → examine query-mix and position changes → report uncertainty and alternative explanations. This studies the combined effect of an editorial change, not NavBoost in isolation.
8 · Step-by-step execution (proposed procedure; not executed)
- Define an authorized page/query cohort and a fixed baseline; select eligible units before inspecting treatment outcomes.
- Improve its real CTR legitimately first: rewrite the title/meta to be more compelling. A title change can affect relevance and presentation directly, so it does not isolate CTR causation.
- Retain comparable untreated pages and document concurrent site changes, search updates, seasonality and shifts in query mix. Do not infer a click mechanism from a before/after association.
This archive does not provide a turnkey click-bot recipe: the reliable, ethical lever is genuinely improving click appeal, and the manipulated version is both legally grey (fraudulent traffic) and empirically shaky.
9 · Demonstration
Supplied practice: Offline Lab 1 demonstrates why aggregate changes do not identify the effect of a manipulated input. Its captured output is distinct from any unexecuted live protocol below.
Hypothetical scenario
An invented example: clicks rise and average position improves in the same week. That observation fits several explanations, including higher positions causing more clicks, a title change improving relevance, or a shift toward branded queries. The observation alone cannot distinguish them. No measurements or expected effect size are implied.
Unexecuted observational protocol. Compare documented title/snippet changes with matched untreated pages over repeated intervals. Measure impressions, clicks, CTR, position and query composition. Randomized page-level assignment can strengthen a study, but interference and title relevance still prevent interpreting it as “CTR alone.” No original outcomes are reported.
10 · Expected output & results
The expected output is a dated table of Search Console metrics, treatment records and uncertainty, including null results. This archive offers no supported estimate of purchased-click gains, duration or filtering rates. Before/after vendor screenshots are insufficient to establish causation.
11 · Documented real-world context
September 2023 testimony in US v. Google discusses NavBoost and click information. The primary record supports system existence and click use; it is not a purchased-click experiment and gives no transferable recipe or efficacy rate. [3]
12 · Common mistakes
Confusing correlation with causation; comparing changing query mixtures; omitting untreated pages; treating “squashing” as a confirmed anti-bot implementation; and accepting an improvement screenshot without the baseline, collection method and later observations.
13 · What experts get wrong
14 · Detection methods
Unexpected click or geography patterns can prompt an investigation, but do not establish fraud. Compare provider logs and analytics with Search Console aggregates, noting their different definitions. Google’s exact production filtering and cross-checking mechanisms are not established by the cited field names.
15 · Defense & recovery
Do not treat anomalous clicks as proof that a competitor altered rankings. Verify analytics configuration, bot filtering and query mix; document any vendor activity. Improve titles and content for their usefulness to searchers while measuring the combined editorial outcome.
16 · Risk & legal status
Buying bot or fraudulent-click traffic can breach the click provider's and Google's terms, and generating fraudulent traffic can carry legal exposure (wire fraud / computer-misuse framings) depending on method and jurisdiction. The bigger practical risk is simply paying for an effect that does not durably materialise.
17 · Expert analysis
Three claims require three kinds of evidence: court testimony can establish that click information is used; documentation can describe stored fields; a well-designed experiment would be needed to evaluate a manipulation’s effect. The first two cannot substitute for the third. The archive’s study is a proposal and provides no ranking-efficacy result.
Field guide: diagnose clicks without attributing a hidden mechanism
Working requirements: a spreadsheet or local script, authorized daily performance exports and a deployment log. Preserve page, query, country, device, date, clicks, impressions and the available position statistic. A basic review needs no paid click service. A larger study needs a reliable export schedule, access controls and enough independent units for its declared analysis. Cost is analyst time plus any existing export infrastructure; no purchase volume or universal study duration is justified here.
The causal problem is feedback. Rank changes which results receive impressions and clicks; a title changes relevance and presentation; changing demand alters the query population. An aggregate CTR increase can therefore arise without a single searcher changing behavior. An observed click series is not a view into the engine’s accepted, filtered or transformed training signals.
Constructed query-mix example
| Fictional query group | Period A | Period B |
|---|---|---|
| Brand query, CTR held at 50% | 100 impressions, 50 clicks | 900 impressions, 450 clicks |
| Nonbrand query, CTR held at 5% | 900 impressions, 45 clicks | 100 impressions, 5 clicks |
| Aggregate | 95 / 1,000 = 9.5% | 455 / 1,000 = 45.5% |
These invented values illustrate composition, not an experiment result. Neither group’s CTR improved, yet the aggregate rose by 36 percentage points. A dashboard that labels this “better engagement” without inspecting the mix would misstate the evidence. Position, country and device mix can require similar stratification; noisy small cells should not be converted into confident conclusions.
- Freeze the query/page cohort and define the comparison before inspecting the post-change outcome. Keep missing or anonymized query coverage visible.
- Reconcile clicks and impressions to the export totals. Flag changing denominators and data gaps rather than filling them with zero.
- Annotate deployments, outages, migrations and known update periods. Their coincidence is context, not proof of causation.
- Compare stable strata and untreated units. Use the same observation windows and record departures from the plan.
- Report both the observed change and the most plausible alternative explanation. If assignment was not randomized and confounding is unresolved, call the finding an association.
Exit condition: a report another analyst can reconstruct from the retained exports and code. Null or contradictory results are valid outcomes. A fixed improvement target would create pressure to stop selectively when a noisy metric looks favorable. The title/snippet protocol in Study 4 remains unexecuted because no real dataset is available.
18 · References
- Leaked Google Search API documents (NavBoost, click attributes). Rand Fishkin, SparkToro, 2024. secondary reporting / analyst interpretation
- Secrets from the Algorithm (NavBoost analysis). Mike King, iPullRank, 2024. interpretation
- US v. Google — public September 2023 trial testimony on NavBoost. primary