The Searchmetric AI Visibility Score
Eight components, four engines, reported separately and never averaged into one flattering number. This is the whole method, including what it cannot tell you.
Published because the decisive question about AEO and GEO is not "will it work" — it is "how would you prove it worked".
We started at two
On 8 August 2026 we set our own baseline for searchmetric.in using a third-party AI-responses index: two responses, two cited pages — one in ChatGPT, one in Perplexity, and zero in Google AI Overviews, AI Mode, Gemini and Copilot.
For an agency selling answer engine optimisation, that is not a comfortable number to publish. We publish it because a methodology page that only shows finished results is a brochure, and because the same instrument read on the same day by a second tool returned a different count — which is itself the most useful thing a prospective client can know about AI visibility measurement.
Every client engagement starts the same way: a dated baseline, recorded before any work, with the instrument named.
Eight things we measure
Each is reported per engine. None of them is combined with another into a headline figure.
Brand mention rate
Of the locked questions, the share where the business is named anywhere in the answer. Named and cited are different things and we never merge them.
Recommendation rate
The share where the business is not just mentioned but actually put forward as an answer. This is the number that correlates with enquiries.
Citation rate
The share where the business's own domain appears as a cited source. In several categories we have measured this at zero while mention rate was healthy.
Competitor share
Which competitors appear, how often, and in what order. A rising competitor share with a flat mention rate is a different problem from a flat market.
Prompt coverage
How much of the locked question set returns the business at all. Coverage tells you whether the problem is one category or the whole entity.
Source authority
Which domains supplied the answer — own site, directory, review platform, forum, news. This decides where the next month's budget goes.
Sentiment
How the business is characterised when named. An unprompted negative aside sourced from a forum is a real finding and we report it.
Answer prominence
Where in the answer the business appears — first named, in a list, or in a trailing caveat. Position inside an answer is not a ranking, but it is not nothing.
Three layers, in this order
Google's own generative-AI reporting
Search Console's generative-AI performance reports are first-party and free. They give impressions in AI features, the URLs that appeared, country and device. They do not give clicks, CTR or query data, and AI Mode referrer data is frequently hidden. We lead with this layer because it is Google's own instrument, and we state its limits in the same breath rather than quietly filling the gaps with estimates.
Manual citation tracking against a frozen question set
Fifteen buyer questions, agreed at kickoff and then locked. Run through ChatGPT, Perplexity, Gemini and Google AI Overviews, same wording, same day each month, screenshotted. This is the layer that is specific to the client rather than to the market, and it is the one that takes real time — about an hour a month. A question set that drifts produces movement that is pure instrumentation, so ours does not drift.
Third-party tooling, as a supplement
AI-visibility panels from the major SEO platforms are useful for share-of-voice and competitor comparison. They are not the proof, and they disagree with each other: on 8 August 2026 two of them reported different citation counts for this very domain on the same day. We report tooling third, and we say which tool produced which number.
Agencies that lead with a third-party tool score are leading with their weakest evidence. The order matters as much as the layers.
What this score is not
An AI visibility score is not equivalent to a Google ranking, and treating it as one will mislead you. The same question, asked twice, can return different businesses. The output moves with the model, the exact phrasing, the location, the user's own context, the date, whether the model browsed the live web, and model updates that are never announced.
So the score is a measured rate across repeated, dated, identically-worded runs — closer to a poll than to a rank. Like a poll, it has a margin of error we do not pretend away, and it is informative in trend rather than in any single reading.
Things we will not do
- Blend four engines into one number, or chart them on one line
- Report a question we did not run as a business that was not cited
- Publish a citation we did not observe, or a prompt we did not run
- Promise a guaranteed position in an AI answer, or a date by which one will appear
- Imply that any of this influences what a model was trained on
The method, applied in public
We run this on our own market and publish the results, with the prompts, the engines, the date and the raw counts, so anyone can contradict us:
- Mumbai, ChatGPT, five categories — zero business websites cited across all five; the review profile did the work.
- Kochi, Perplexity and AI Overviews, twelve categories — 18 of 46 cited sources were businesses' own sites, split sharply by category.
- Delhi, paired-query test — every commercial query returned no AI Overview; every advice-framed one did.
If your business is absent from these answers and you want the reasoning rather than the measurement, start with why your business isn't showing on ChatGPT.
Measuring AI visibility — your questions
Is the AI Visibility Score a single number?
No, and this is the most important thing about it. Roughly 11% of the domains ChatGPT cites are also cited by Perplexity, so the engines are largely measuring different worlds. We report four engines as four separate lines and eight components within each. Any agency handing you one blended AI visibility number is handing you a figure that describes nothing in particular.
Is an AI mention the same as a Google ranking?
No. A Google ranking is a stable, reproducible position for a query in a location. An AI answer varies by model, by how the prompt is phrased, by location, by the user's own context and history, by date, by whether the model browsed, and by model updates you will never be told about. We report AI visibility as a measured rate across repeated runs, never as a position, and we do not convert it into one.
How do you handle a question you have not checked this month?
It is logged as not checked, not as not cited. Those two are different findings and conflating them corrupts the series — we rewrote one of our own early run logs for exactly that mistake. An honest gap in the data is more useful to you than a confident zero.
Can you guarantee we will appear in ChatGPT?
No, and you should treat any agency that does with suspicion. Published research puts brand-citation rates in ChatGPT responses very low, and a large share of the pages it cites most are structurally unreachable for a business — Wikipedia, government and university domains, app stores, major news. What we can commit to is monthly measurement across four engines with screenshots, the specific structural work, and telling you when something is not working.
How long before the numbers move?
Perplexity responds fastest because it weights freshness heavily and re-retrieves constantly, so changes can show within weeks of indexing. Google AI Overviews typically take one to three months, and only where an Overview exists for the query at all. ChatGPT is slowest, leaning on a corpus and third-party corroboration, and is better measured over two or three quarters than month to month.