Insights Data integrity

We don't dress up the numbers

GEO makes it easy to produce good-looking numbers. This page lists four rules already enforced in the product; each one can make the numbers look worse.

Four principles

Four rules we chose to follow

The gray line under each principle is what we could have done instead — easier, better-looking, and unable to survive a client's review.

Zero samples read "no data", never 0%

When an engine yielded no valid samples this period, the report says "no valid sampling this period" — not 0%. "Not measured" and "measured at zero" are opposite findings: one is missing data, the other is a real negative result.

The easier route: print 0% — tidy tables, pretty charts, and clients misreading "this engine never mentions us" when the truth is we never measured it.

Every metric carries its denominator and sample size

Every external number states its numerator and denominator (e.g. mention rate = mentioning samples / valid samples) plus the sample size. Insufficient samples are labeled as such, with no conclusion drawn.

The easier route: hand over a single percentage — unverifiable, incomparable, and impossible to explain when questioned.

Data from different capture methods never merges into one trend line

Official API, UI simulation, manual entry and third-party sources are recorded and presented separately. Averaging two methods produces a line that looks perfectly normal and does not exist in reality.

The easier route: draw one smooth line mixing everything — visually "complete", and describing a trend that never happened.

A "false improvement" self-check: every change must be explainable

Re-tests disclose a measurement-condition snapshot (question set, platform set, competitor set, sampling window). When a model or question-set change moves the numbers, the report marks the periods as not directly comparable — measurement changes are never repackaged as optimization wins.

The easier route: hide the change — up means "optimization worked", down means "model fluctuation".

Capture methods and reproducibility

Every data source states what it can prove — and what it cannot

This table shares its source of truth with the appendix in client reports (one set of definitions) — the public page and the deliverable cannot disagree.

Capture method	Reproducible	What it proves	What it cannot prove
Official API	Reproducible	We made a real request to the engine official interface and retained the full answer; the same input reproduces the same output.	It does not prove that ordinary users see the same answer in the app or web product - the API and the consumer product may differ in retrieval and personalization.
Licensed third-party source	Not reproducible	Results were obtained through a licensed channel, with its definition and capture time stated.	It does not prove the third-party definition matches our own sampling - figures from different definitions must not be compared or merged.
UI simulation (browser)	Not reproducible	A real browser walked the same path a user would, covering engines with no public API.	Not reproducible: page structure, gradual rollouts and login state all change results, and identical answers are not guaranteed.
Manual entry	Not reproducible	A person tested the engine directly and recorded the result, for cases that cannot be automated or as a cross-check.	Not reproducible: wording, account state and timing cannot be recreated exactly, so it should not drive a trend line.

Reproducible = re-running under the same conditions yields consistent results. Non-reproducible methods (UI simulation, which shifts with page redesigns) are labeled separately and never merged into other sources.

Known limits

4. Known limits of this method

No ranking on small samples
When the source matrix has fewer than 10 samples we show a notice instead of a ranking - a top-source list from a handful of samples is noise, and acting on it wastes budget.
Capture methods are never merged
Official API, third-party, UI simulation and manual entry differ in reproducibility and are never merged into one trend line. Averaging them produces a blended metric that does not exist - and looks perfectly normal.
Brand-named questions are pooled separately
When the question itself contains the brand (e.g. "how is XX"), AI almost always repeats it. Those samples are excluded from the mention-rate numerator and denominator to prevent false positives.
Uncovered markets are not filled in
A market without local-language sampling is shown as "not covered" rather than filled with English samples, which would systematically overstate your visibility there.
Egress region affects results
Reaching the same engine from a domestic versus an overseas egress can return different retrieval results (regional ranking). Samples from the two egresses are not merged; reports label them separately.
Trends beat single points
A single sample fluctuates by nature: in field measurements, re-running the same prompts on the same day yields a source-overlap of only 0.32–0.43 — a single run carries almost no signal. Judge by trends over consecutive weeks, and note that a rise or fall across a capture-method switch may come from the switch, not from brand performance.
Engine-side sampling parameters are beyond our control
We cannot control the engines' internal sampling parameters: peer-reviewed research measured 9–28% decision changes across repeated runs even at temperature 0. Any single snapshot should therefore be read as a distribution — which is exactly why this report uses repeated sampling and confidence intervals.
Numbers are not directly comparable across engines (or across app/web)
Engines differ in retrieval sources and training corpora (research measured systematic gaps of 30+ percentage points in brand mention rates between Chinese and international engines), and the App and web surfaces of the same platform use different source pools. Compare engines by trend only, never as a cross-engine ranking; this report labels each engine separately.
Content quality scores do not predict citation probability
A large-scale study of Chinese-language citations found that a page's "composite quality score" is not a leading predictor of any citation outcome. This platform's article GEO score is a content-engineering aid; it is not a promise that raising the score raises citation probability.
Individual AI answers fluctuate by nature (this is not a decline)
The AI re-runs retrieval and cross-validation each time: the same question can yield different answers and different cited sources tomorrow — external research measured roughly 65% of cited sources changing between adjacent days. So "we were mentioned last week but not this week" is usually mechanical volatility; judge the real trend from overall rates across many sampling rounds (this platform uses multi-round sampling and confidence intervals).
Transactional and knowledge questions follow different citation rules
For transactional questions (local services, e-commerce, product selection), some engines answer directly from their own ecosystem — merchant databases, product cards, map listings — without citing external pages. For knowledge-oriented questions (industry expertise, B2B proposals, brand reputation), external sources remain the main basis. Citation counts therefore cannot be compared across question types: fewer citations on one type may simply mean that scenario runs in a closed ecosystem, not that your content is weak. This platform reports by question type and does not merge across types.
Search ranking is not the same as being cited by AI
Large-scale research shows about 88.46% of ChatGPT citation URLs come through its search retrieval path (you must be indexed before you can be cited), yet for Google AI Overview only about 37.9% of citations come from the top 10 organic results — the rest come from positions 11-100 and beyond 100 (Ahrefs, March 2026, analysis of 4 million URLs; the same firm reported about 76% in 2025, more than doubling within about a year, so such ratios must be read with their timestamp). Being ranked #1 is thus neither necessary nor sufficient for citation. This platform counts only citations that actually appear in AI answers and does not use search ranking as a proxy.

What this cannot tell you

5. What these numbers cannot tell you

API results are not what consumers see (the API and the consumer product can differ in strategy)
Not appearing this time is not the same as not existing (we may simply not have sampled it yet)
Being cited by AI is not conversion (landing pages, pricing and sales sit between the mention and the order)
A rising trend is not proof it was our doing (competitor moves and engine updates also change results)
Figures are not comparable across markets (the same "visibility rate" means different things, so read them with their definition)

These limits and inference restrictions appear verbatim in the appendix of every client report — not just as a posture on this website. The full methodology lives in the measurement methodology page.

Judge the credibility of the data before the data itself

Run a free AI visibility check Methodology