An AI visibility report is built not around a single number but around a fixed set of control queries and seven metrics: mention rate, share of voice, recommendation rate, citation rate, average position in a list, factual accuracy and traffic from AI services. They have to be counted every month against the same query set, otherwise the months cannot be compared.
Key takeaways
A single number answers none of the working questions. “AI visibility 64” says nothing about whether you are recommended, whether the site is linked, or whether your services are described correctly.
A useful report answers a chain of questions in order: does the model know the company, does it name it in an answer, does it recommend it, where does it stand among competitors, does it link to the site, which sources it uses instead of you, how it describes the company, how accurate that description is, and whether any of it brings people and enquiries.
Every question in the chain is a separate metric. Rolling them into one score is fine for a title slide, but decisions cannot be made on that score: it hides a rise in mentions alongside a fall in links just as easily as the reverse.
Measurement starts with the list of questions a prospective client asks, not with a tool. The list is written down and does not change without reason — it is what makes the numbers comparable.
For an IT company in Tashkent the set looks like this: CRM development in Tashkent, business automation companies in Uzbekistan, who builds ERP for distributors, Telegram bot development, AI for a sales team, how much CRM development costs, alternatives to a specific vendor, best IT companies in Uzbekistan.
The size of the set matters more than its cleverness. Twenty queries checked every month are worth more than two hundred checked once.
The second rule is to write the query down word for word, along with its language and region. A question rephrased on the fly returns a different answer, and a month later you will be comparing two separate checks rather than a trend.
There are seven metrics, and each answers its own question in the chain. The formulas are simple: every one is a share of a clearly named denominator.
| Metric | Question it answers | How it is calculated | Example |
|---|---|---|---|
| Mention rate | Does the model know the company | Answers mentioning the brand divided by the number of queries checked | 42 mentions across 100 queries |
| Share of voice | How much of the category you hold | Your mentions divided by the mentions of all brands in the same answers | 42 out of 300 mentions of all companies |
| Recommendation rate | Are you named or actually recommended | Answers with a recommendation divided by answers with a mention | 32 recommendations across 50 mentions |
| Citation rate | Does the answer lead to your site | Answers linking to your domain divided by answers with a mention | 25 links across 40 answers with the brand |
| Average position in a list | Where you stand in a list of companies | The brand's average place in enumerations | third on average across 20 answers |
| Factual accuracy | Are you described correctly | Correct statements about services, prices and geography divided by all statements checked | 18 correct statements out of 20 |
| Traffic and enquiries | Does any of it bring people | Sessions, enquiries and sales from AI service sources | sessions from ChatGPT over a month |
The cadence differs. Weekly you look at mentions for the main commercial queries and at traffic. Monthly you count all seven metrics. Quarterly you work through which sources the models use instead of you and what to do about it.
Average position in a list is the one metric that must not be read as a search position. The list inside an answer is built around the exact wording of the question and shifts with it, so it is read only as a trend and only against the same competitors.
Some of the numbers come from the platforms themselves, and that is where to start: it is free and needs no manual work. The rest is collected by hand against the control queries.
| Source | What it shows | What it does not show |
|---|---|---|
| Google Search Console, generative AI performance report | Impressions of your pages in AI Overviews and AI Mode, by page, country and device (Search Console help) | No clicks or positions in the report; not available to every site |
| Bing Webmaster Tools, AI performance report | How often the site is cited in Copilot answers and Bing summaries, which pages and for which queries (Bing blog) | It counts citations, not visits |
| Bing Webmaster Tools, citation share | Your share of citations among all citations shown for the same query (Bing blog) | Microsoft surfaces only |
| Yandex Webmaster, visibility in Alisa answers | The share of queries that mention the site, plus sample queries, pages and competitors (Yandex Webmaster help) | Only Alisa answers in Yandex search |
| Manual checks against the control queries | ChatGPT, Perplexity, Claude and anything else that gives site owners no reports | Labour-intensive, and repeats are needed because answers are unstable |
Traffic is counted in web analytics. Google Analytics 4 now has a dedicated channel for it: visits from ChatGPT, Gemini, Copilot, Grok and DeepSeek land there automatically, with the source marked ai-assistant (Google Analytics help). One caveat matters: visits from AI Overviews and AI Mode do not go into that channel — they arrive as ordinary search and are only visible in Search Console.
Because model outputs are non-deterministic. OpenAI puts it plainly: by default the model may return a different result for the same request, and the parameter meant to make output repeatable produces “mostly” the same answer without guaranteeing it (OpenAI documentation).
Hence the rule of measurement: a single check proves nothing. The same query is run several times, the result is averaged, and what goes into the report is the trend rather than one answer.
Note Record the conditions of the measurement alongside the figures: date, service and model version, query language, region, number of repeats, and whether the search mode was on. Without those fields, comparing months turns into an argument about method.
The report reads top to bottom and ends with tasks rather than charts. The order of sections stays the same month after month, which is why people get through it.
If the report does not fit on five pages, the problem is not the volume of data but the absence of conclusions.
Every conclusion is carried through to six fields, otherwise it is an observation rather than a task. The format is the one we use in our article on turning analytics results into an action plan: one format across the company saves reading time.
| Field | What goes in it | Example from a report |
|---|---|---|
| Fact | What the measurement showed and for which period | Zero mentions for “ERP for distributors” across three months |
| Cause | Why it is happening | The site has no page that answers this question in full |
| Action | What exactly changes | Write an article covering the task and the order of implementation |
| Expected effect | Which metric should move and where | Mentions appearing for this query in at least some answers |
| Owner | One person, not a department | The blog editor |
| Deadline | The date the change has to be working | By the next monthly measurement |
Factual errors deserve a line of their own. When models credit the company with services it does not offer or quote outdated prices, the fix is publishing unambiguous facts on your own site and updating directory listings, not writing to the service's support desk.
An SEO report answers the question of where the site stands. An AI visibility report answers what is said about the company and what is linked while saying it. The first counts impressions, clicks and positions; the second counts mentions, recommendations and citations.
The two do not compete and rest on each other: a page that is not indexed will not reach an AI answer either. How the foundation works is covered in our article on whether you still need SEO in the age of AI, and what to do with pages for the sake of answers is in the article on generative engine optimisation.
The last link in the chain is money. Until visits from AI services are tied to enquiries and deals, the report stays a report about visibility. How to assemble the path from source to payment into one chain is shown in our article on end-to-end analytics.
Syntra Systems does not sell a “guarantee of mentions” — the platforms do not publish their source selection rules, so there is nothing to promise. We build the control query set with you, set up the reports in Search Console, Bing Webmaster Tools, Yandex Webmaster and web analytics, run manual checks for the services that give no reports, and hand over a monthly analysis with an action plan.
This work sits inside project development, from $1,500 a month: a fixed number of team hours, sprints with a demonstration of the result and a quarterly review of priorities. The terms are described on the technical support page.
Let’s discuss your project
Tell us what you need, and we will estimate the timeline and cost and suggest a solution.
With twenty control queries and one service. Check them by hand, record the share of answers that mention the brand and the share that link to the site. That is the first baseline; from there the set grows rather than being rewritten.
Not at the start. The free reports in Search Console, Bing Webmaster Tools and Yandex Webmaster, plus a manual check against the query list, cover the basic picture. A paid tool makes sense once there are more than a hundred queries and manual checking eats a working day.
A mention is when the company is simply named in a list. A recommendation is when the model explains why it is worth approaching or puts it first with a reason. Counting the two as one metric is wrong: they mean different things.
Find where the wrong data comes from: usually an outdated page on your own site, an old directory listing or someone else's retelling. Fix the original, publish an unambiguous statement of your own, then repeat the measurement a month later.
Through web analytics and the CRM: visits from AI services are marked as a separate channel, and from there the chain runs to enquiries and deals. Until that chain exists, the report shows visibility rather than money.
No more than once a quarter, and always with a note in the report. New queries can be added, but old ones are better kept: they are what makes comparison with previous months possible. Replacing the set wholesale wipes out the accumulated history.
Sources
