Does ChatGPT name you when your buyer asks?

Every analytics tool you own, Metrivo included, is a record of what happened on your website. It sees a visit, a session, a page, a payment. It is structurally blind to the moment that matters most in an AI-mediated purchase: a buyer typed a question into ChatGPT, read an answer that named three products, and your product was not one of them. Nobody visited. Nothing was logged. The loss is invisible.

This is a different failure from the one most founders are worried about. The familiar problem is that AI-referred traffic arrives without a referrer and lands in the direct bucket, which is a measurement problem you can partly solve with better instrumentation. The problem on this page is upstream of traffic entirely: you were never in the answer, so there was never a visit to measure. No amount of attribution work recovers a customer who was recommended a competitor.

The only way to observe it is to ask the engines the same questions your buyers ask, repeatedly, and record what comes back. That is what a visibility scan is. It is not a ranking check, it is not a crawl, and it is not an estimate derived from your traffic. It is a stored observation of what a model said at a moment in time, which is a genuinely different kind of data from anything else in your stack.

You describe your brand once: the product name, the category it competes in, a short description, and the competitor domains you want tracked. Metrivo can suggest an initial set of buyer prompts from that description, but you own the final list. This matters more than it sounds. A prompt written to make you look good is worthless. The useful prompts are the ones a buyer would actually type without knowing you exist, phrased as a category problem rather than a brand name.

A run then sends each prompt to each controlled engine through Metrivo's server-side provider. Running server-side rather than in your browser is deliberate: a logged-in ChatGPT session carries memory, custom instructions, and personalisation, all of which can quietly bias an answer toward a product you have already been reading about. Your own account is the worst possible instrument for measuring whether strangers get recommended your product.

Each response is parsed into citations. A citation records the brand named, the domain it maps to, the URL cited where the engine gave one, the position it appeared in, and a sentiment label of recommended, mentioned, neutral, or negative. That sentiment split matters: being listed as an also-ran in position seven is not the same outcome as being the engine's first recommendation, and a visibility metric that treats them identically will tell you that you are winning while you lose.

Runs are idempotent and every observation keeps its engine, model identifier, prompt snapshot, response hash, and timestamp. When a scan cannot complete, because an engine is unreachable or a response cannot be parsed, Metrivo records the failure rather than silently dropping the prompt. A visibility trend with invisible gaps in it is worse than no trend.

The first scan is almost always uncomfortable, and the most common reaction is the wrong one. Founders see that they were named in two prompts out of ten and immediately start writing content. Before that, check the crawler log. If GPTBot and PerplexityBot have never fetched the pages that would answer those prompts, content is not your constraint; accessibility is. Check robots.txt, check that the answer actually exists on a server-rendered page rather than behind JavaScript or a signup wall, and re-scan after the crawlers have had time to return.

The second thing to check is which competitors were named instead, and read those answers closely. Engines tend to cite pages that state a specific, checkable claim about a specific job. If a competitor is being named for a capability you also have, the gap is usually that you never wrote the sentence that says so plainly on a page an engine can read.

Third, watch sentiment and position over several runs rather than reacting to one. Movement from "mentioned in position six" to "recommended in position two" on a high-intent prompt is a real signal. A single answer changing between runs is often just model variance. This is why the trend matters more than any individual scan, and why we store every run instead of overwriting a current score.

Finally, resist the temptation to optimise for the prompts you are already winning. The valuable ones are the high-intent questions where a buyer is choosing between named options and you are absent. Those are the answers that lose you revenue silently.

Being recommended is worth nothing if the resulting visitors bounce off your pricing page, and converting AI traffic beautifully is worth little if you are named in one answer a month. The two measurements only become a strategy when you read them together, which is why Metrivo runs AI-search attribution alongside visibility rather than selling them as separate products.

The honest limit sits between them. When an engine names you and the buyer copies your URL into a new tab, the visit typically arrives with no referrer, and Metrivo will not relabel that as AI revenue. See AI traffic detection for exactly which signals are treated as confirmed, inferred, or unknown, and AI crawler tracking for how crawler visits are separated from human sessions.