Version 1.1, updated 15 Oct 2026
How we measure AI answer visibility, and why we publish a range
Answer engines are non-deterministic. The same question asked twice can return different brands. A single decimal-place score hides that; a range reports it.
What counts as a mention
A mention is recorded when the brand name, a registered product name or an unambiguous paraphrase appears in the answer body of a sampled response. Mentions inside a citation title only are recorded separately as a source appearance, not a mention.
Five verdicts, not one score
Each answer is scored as Top pick (named first or called the best), Recommended (suggested as a good choice), Mentioned in passing (named, but not backed), Named with caution (named, but with a warning attached) or Not mentioned. A brand named once in a list is never counted the same as a brand an engine recommends. Every verdict links to the exact sentence the engine wrote and the rule that scored it.
Discovery and recognition, never averaged
Questions are asked in pairs. Unbranded questions test whether an engine finds the brand without being told its name. Branded questions test whether it knows what the brand is. Discovery, recognition and mention rate are reported as separate figures, each with its sample size, and are never blended into one number.
Answers from an engine's built-in knowledge (Trained) and from live web search (Live) are tracked and charted separately. Mention rate is also broken down by question type, such as open recommendation, direct comparison and how-to-choose, because the same brand often wins one and loses another. All questions are asked in English; the market setting controls location only.
Comparisons and watch items
The brand is compared only with the competitors it tracks. We never invent an industry average. Answers that confuse the brand with something else, or place it in the wrong category, become watch items that grow more prominent the longer they stay unresolved.
Checking whether a fix worked
When an action is marked complete, we compare the four weeks before it with the two and four weeks after. A win or a drop is only reported when the change is bigger than normal week-to-week variation. Drafted pages leave any fact we do not know as a clear [ADD: …] placeholder, never an invented one.
Engines sampled
ChatGPT · Perplexity · Google AI Overviews · Gemini · Copilot · Claude
Engines are sampled on a rolling schedule rather than all at once, so results are not skewed by a single day's model behaviour.
How each engine is captured
Perplexity and Claude are sampled through their official APIs. ChatGPT and Gemini are sampled through model APIs rather than the consumer apps, so answers carry no personalisation or memory; the exact model version is pinned and recorded with every response, and a model change is treated as a methodology change, not a score change.
Google AI Overviews and Microsoft Copilot have no official API. They are captured from the live search surface through an independent search-data provider, and each needs a provider that returns that engine's answer block specifically. The provider returns the AI Overview and the Copilot answer as structured data (answer text, cited sources and timestamp). A second independent provider is retained as a backup datafeed so a capture gap on one feed never silences an engine. For Copilot, access to general Bing search results is not the same thing: the provider must return the Copilot answer itself, which only specialist endpoints offer. An answer is only recorded when the engine actually serves one; we never manufacture or infer one. Where an engine is captured from the search surface rather than a model API, samples are marked accordingly, because they reflect what a user sees on screen rather than a clean model output.
How the range is produced
For each prompt we sample multiple responses per engine across the reporting window. The lower bound is the share of answers mentioning the brand at the 10th percentile of sampled days; the upper bound is the 90th. The band width is therefore a measure of instability, not of uncertainty in our data collection.
How confidence is calculated
Confidence is high when the band is narrow, sample volume is sufficient and engines agree. It is low when a single volatile source, often one forum thread, accounts for a large share of mentions. We do not report a trend direction until the change exceeds the band width across two consecutive windows.
How the Rankings table works
Rankings use only answers to questions that do not name the brand, so every mention is unprompted. Answers from trained knowledge and from live search are ranked separately and never mixed. A table is only put in order once it has at least 20 answers (one engine's reply to one question counts as one).
Mention share is the share of answers that name a company positively, an answer counts once, however often the name repeats, and criticism or "alternatives to…" wording does not count. Top pick and average list position come only from answers that lay out a clear ranking: a numbered list starting at 1, or "Best overall / Best for" labels. The order names happen to appear in ordinary sentences is never used.
Different spellings of one company (with or without a web address, a leading "The", or source markers like [1]) are grouped into one row. Headings, product types and advice lines are not treated as companies. The brand's own nicknames fold into its row, which is always shown and highlighted.
Radical transparency: every caveat in our data, stated in full
No measurement of AI answer engines is perfect. This section lists every limitation and trade-off in how Seenford captures data, so you can weigh our numbers with your eyes open.
Current engine status
All six engines are live. ChatGPT, Gemini, Claude and Perplexity answer directly at source, while Google AI Overviews and Copilot are captured through a third-party data provider.
A clean session, not a personalised one
Our captures have no account, no search history and no cookies, the equivalent of an incognito window on a UK desktop. That is a deliberate choice: it is consistent and reproducible, which is what trend tracking needs. But a signed-in user with a search history may see a slightly different answer, and we cannot, and should not, try to replicate that.
Answers vary even for the same user
Engines decide per session whether to serve an AI-generated answer at all, and regenerate the text frequently. There is no single "the" answer to match — what we capture is a genuine instance of what a real user would have seen at that moment, not a canonical version. This is why we report ranges, not single scores.
Location and device are set, not observed
Every capture is pinned to a country and a device you choose (United Kingdom on desktop by default), encoded into Google's own location parameter so the answer matches what a searcher in that country would see. A user on a different device, or in a specific city, may see something different. Each country and device is treated as a separate sample and labelled as such, results from different markets are never blended into one number.
During setup we may suggest a country based on where your device appears to be, a starting point you can accept or change. It is only ever a suggestion: it is never applied without your confirmation, and it is never re-applied automatically on later sessions. Your tracking market does not silently drift because of where you happened to sign in from.
Trained knowledge or a live look-up, two different measurements
When you ask ChatGPT about a brand today, it often searches the web mid-answer. When we ask the same question through the model's interface with nothing else switched on, it answers from what it learned during training. Those are two different things, and they can say the opposite of each other on the same day.
A real example from our own testing. We asked an engine what it knew about a small Reddit-intelligence product. Our capture, trained knowledge only, replied that it did not have reliable information about the company. The same engine in a consumer app, minutes later, produced an accurate profile, naming the product's own published reports and a brand comparison it runs, because it had just read the company's website while answering. Neither answer is wrong. One measures whether the model knows you; the other measures whether an agent sent to look you up can piece you together from what it finds.
Live data is our default and is the closest reproducible approximation of a fresh, live user. It is also the metric you can actually move: a live answer is built from what is on the web right now, your site, your comparison pages, your coverage, which is the thing you can change.
You choose the mode per brand in Settings, and live is the starting point. We never merge them: every stored answer carries its mode, scores are calculated within a mode, and a chart never joins a trained-knowledge point to a live one. ChatGPT, Gemini and Claude all run in both modes, the engine's own web search is switched on in live mode. Perplexity, Google AI Overviews and Copilot are live search surfaces by nature — the setting does not apply to them, and they are always labelled as live.
Asking a question does not change what the engines say
A common worry: does querying these engines on your brand's behalf change what they tell other people? No. None of these systems learn from a query in real time. Models are trained in separate, infrequent runs on massive datasets, a single API call (or the thousands we make) never retrains or fine-tunes anything live. Even if a query were eventually included in some future training batch, one mention among billions of tokens does not move what a model says about your brand. There is no mechanism by which asking a question today makes an engine more likely to recommend you tomorrow.
Whether a query could reach a future training set at all depends on each provider's data policy, and ours are the strict end of it:
- ChatGPT (OpenAI), our API account is set to exclude inputs and outputs from training entirely.
- Claude (Anthropic), commercial API usage is not used to train models by default under their terms.
- Gemini (Google), the paid API tier is not used for model improvement by default (unlike the consumer app, whose policies differ, we do not touch it).
- Perplexity, AI Overviews and Copilot, these are live search and retrieval systems reading the current web, not training-feedback loops; querying them does not adjust any underlying model.
This also matters for privacy: because your brand questions travel over API tiers that exclude training use, the names, markets and competitor details you enter are not fed into anyone's models, they are processed to answer your question and go no further. It is the same principle we apply to personal data under GDPR: collected for a stated purpose, processed for that purpose only.
An answer is only recorded when the engine serves one
Google does not serve an AI Overview for every query. When none is served, we record "not served", we never manufacture or infer one. Absence is data, and we report it as such.
Cost transparency
Every live query has a real cost. Sampling volume is a cost decision as much as a statistical one, and we balance that to ensure maximum knowledge retrieval per brand, per month.
What we do not claim
We do not measure how many people saw an answer, and we do not attribute revenue to AI visibility. We do not report a trend direction until a change exceeds the band width across two consecutive windows. Anything we cannot evidence with a prompt, a response excerpt and a timestamp is not reported.
See it on a real brand
Run a Free Report to see these rules applied to one brand across six engines, compare the plans, or read 50 questions about AI visibility.