AI visibility

​

What this shows

The AI Visibility page answers one question: when someone asks an AI assistant about your category, do you come up? Traceten asks your prompts across answer engines on a schedule, stores the answers, and measures three separate things:

• Presence rate: how often you are named
• Citation rate: how often a page you own is cited
• Share of voice: your mentions as a proportion of every tracked brand’s mentions

They are never blended into one score. You can be named without being cited (the model learned you) and cited without being named (it used your page and credited nobody), and averaging those produces a number that cannot be acted on. The exact formulas are on How AI visibility is measured. This is a different signal from AI sources and AI crawlers. AI sources measures the person who arrived from an assistant. AI crawlers measures the fetch that put you in its answer. AI visibility measures what the assistant actually said.

​

What this is not

These numbers come from a poll of the prompt set you wrote. They are not an observation of what real people asked. Real prompts are not recoverable. When someone arrives from an AI assistant, what reaches us is the referring origin or a utm_source value: utm_source=chatgpt.com names the assistant and carries no question. Nobody selling this category of product is measuring real user prompts, including us. So every figure here reads as: of the questions you chose to ask, this is how often you were named and cited. How useful that is depends on whether your prompt set resembles what your buyers actually ask, which is your judgement to make. The two measurements on your dashboard that are unsampled observations are AI crawlers and AI sources, and the Citations tab below joins this page to both.

​

The engines we sample

The dashboard names each engine by the product you know: ChatGPT, Claude, Perplexity and Google AI Overviews. It does not show which model answered or how we reached the engine, because models change often and neither detail helps you read an answer. This table is where that is written down. The distinction is not cosmetic. Google AI Overviews is a results page a real person would have seen, and Perplexity’s Sonar is the retrieval stack behind its own product. Claude and ChatGPT, though, are measured through each vendor’s API, which is not the app. The apps add memory and personalisation before they retrieve, so they can name different products and cite different pages. The visibility API returns the exact engine label and model behind every stored answer, so any number can still be traced. AI answer engines covers each surface in full, including what we send to each provider and what we store from the answer.

​

Adding or removing engine coverage

Coverage applies from your next scan window, not immediately: a scan records the engines it will use when its window opens, so the figures inside one window are always comparable with each other. Removing coverage works the other way: the premium engines stop being called as soon as the change takes effect, without waiting for the window to end. A call already in flight finishes. That window then finishes labelled as a scan whose coverage ended part-way through, rather than pretending the engines ran.

​

Every number is a range

A presence rate is an estimate from a sample, so the page always shows the 95% confidence interval beside it: as a shaded band on the chart, and as a range under each figure. This matters more than it sounds. At five runs, a rate of 20% has an interval spanning roughly 4% to 62%. Quoting “20%” without that range is a claim the data does not support. A change is only styled as a change when the intervals do not overlap. A move from 3-of-10 to 4-of-10 looks like a 33% improvement and is noise, so the page renders it flat, with the band visible. A dashboard that flags noise is one nobody reads by the time something real happens. The interval is a Wilson score interval over the repetitions behind each prompt, and How AI visibility is measured explains why that method rather than the textbook one.

​

The x-axis is the scan window, not the day

A scan trickles across its cadence window rather than firing every call at once. Plotting the result by day would draw a sawtooth of whichever days a trickle happened to touch, and read as violent week-to-week swings. So each point on the chart is one scan window (“the 10 days ending Sept 2”), and the bucket boundaries come from the scans that actually ran, not from a fixed calendar.

​

Score availability by plan

Scores are gated on sample size, and your plan sets how large a sample you get. When a score is hidden, the page shows one line saying which of the two is the reason, with the full explanation behind the info mark beside it. If your plan covers the trend but too few of your prompts have been scanned yet, the trend stays hidden and the page tells you how many scored runs you have and how many it needs. Gating on statistical validity is deliberate: we would rather show you nothing than a percentage that moves on noise.

​

Citations

The Citations tab shows which URLs the engines linked to, yours and everyone else’s, and joins your own cited pages to what actually happened on them. Four tiles at the top describe your own cited pages:

• Pages cited: pages of yours an engine linked to at least once in the range. Being linked is not the same as being named.
• Times linked: every time one of your pages was linked in an answer. These are event counts, so they add up. The line under the figure gives the busiest day’s engine count on one page.
• Answer fetches: times an AI assistant fetched your cited pages to answer someone’s live question, from the AI crawlers data. Observed and unsampled, and the closest observable proxy for a real prompt. Matched on the page path only, because the crawl data stores no hostname, so URLs sharing a path are counted once in this tile and the line under it says how many did.
• AI landing revenue: revenue from AI sessions that landed on a cited page. A deterministic join over measured sessions and measured orders, not a model. Always in USD, whatever your reporting currency.

Who gets linked compares you with everyone else. A radial splits the links by domain, with all of your URLs as one “Yours” slice and your share in the centre. The largest domains get their own slice and the rest are grouped as “Other domains”. Beside it, ranked bars list the most-linked pages. Show all lists the rest inside the same card, which keeps its height and scrolls. The shares are taken from the 50 most-linked URLs in the range, not from every URL ever linked. Cited pages lists your own pages with their times linked, answer fetches, AI visits and revenue in USD. AI visits are AI sessions that viewed the page, which is a different population from revenue, which counts landings. Two pages sharing a path show the same answer fetch count, so do not add that column up. Select a page to open its detail card:

• At the top: times linked, answer fetches, AI visits and AI revenue in USD
• In answers: days linked, and the busiest day’s engine and prompt counts
• Answer fetches: the fetch count, how many AI companies made them, and View AI crawls, which opens AI crawlers filtered to that page
• Traffic: AI visits, conversions, AI and all landings, and all landing revenue in USD

A page shows 0 answer fetches when no verified AI assistant fetched that exact path in the range. Crawls of a different path for the same content, such as a redirect target or a URL with a trailing slash, count under that other path. Crawls that fail verification are left out, so this count can be lower than the page’s total on AI crawlers. Those are the two kinds of evidence this feature produces: an observation, and a deterministic join over observations. There is no modelled number anywhere on this page. Traceten does not publish a revenue-per-prompt figure and does not allocate revenue across the prompts that cited a page, because doing that needs a model, and a modelled figure is worth less than the confusion it causes when it is read as a measurement. A per-engine citation breakdown is not available. The citation rollup records how many distinct engines cited a URL on its busiest day, not which ones. The page shows it as the busiest day’s engine count (“Busiest day: engines” in a page’s detail) and, being a per-day distinct count, never sums it across days.

​

Prompts and the raw answers

The Prompts tab lists every question we ask on your behalf. A slim meter at the top shows how much of your prompt allowance is used, and turns the warning colour when it is full. At the limit, Add prompt and Suggest are disabled and the line under the meter says why. Each row shows the prompt’s question type, topic and locale, with a Paused badge when it is paused. When your plan includes per-prompt results, the row also shows:

• Named: how often your brand appears in answers to that prompt, drawn as a tick on its 95% interval band, with the rate beside it. Blank means not measured in this range, never zero.
• Runs: scored answers that named you, out of all scored answers.
• Named more: marks for brands named more often than you. When nobody was named more often, faded marks show the brands that were also named.
• Linked instead: marks for sites linked in these answers instead of yours.

Hover or tap a set of marks to list every name with its count. When you have archived prompts, a Live and Archived switch lists them separately. Select a row to open its raw answers. The Prompt actions menu on each row has See answers, Edit details, Change wording, Pause or Resume, and Archive. The answers card is titled with the prompt, with its question type, topic and locale under it. When your plan includes per-prompt results, the card leads with that prompt’s figures: how often it named you with the 95% interval, answers naming you out of all scored answers, brands named more often than you, and sites linked instead. When more than one engine answered, a filter above the answers narrows them to one engine. Traceten keeps each engine’s answer verbatim, so you can read exactly what it said. Each answer shows:

• The engine by name, when it ran and the locale
• The brands it named, with yours marked and the best rank in a list where there is one
• The verbatim text the engine returned, with each brand mention highlighted
• Every source the engine returned, with the ones on a domain you own marked as yours

The 10 most recent answers in the range are shown. Every number on this page decomposes to that text. It is the only real answer to “I don’t believe this number”.

​

Brands

The Brands tab decides what gets measured. Nothing on the Visibility tab is scored until one brand is confirmed as your own, and until then the tab opens with a notice saying so. Every live brand is listed once, in the section for whose it is:

• Your brand: brands you own. The first confirmed one in this list is the brand every rate on the Visibility tab describes. If more than one is marked as yours, all of them are matched, and the tab names the one reported as you.
• Competitors: scored against the same answers as you, so share of voice compares like with like. One answer naming three brands counts for all three.
• Found on your site: brands Traceten proposed from your own pages.

Each site can have up to 25 live brands. Archived brands do not count toward that limit.

​

Brand status

Each row shows its status as a badge. To start measuring a brand we found, select Confirm on its row. The row menu also has Mark as mine and Mark as competitor, which move the brand into that section and confirm it at the same time. A brand we found and you confirm is measured, but only a brand marked as yours is reported as you.

​

Alias warnings

When you add or edit a brand, you can list other names it goes by and the domains it owns. The name, every alias and every domain are terms the matcher looks for, and a citation of a page on one of those domains counts as a citation of that brand. Duplicates, ignoring case, are counted once. Some terms are saved but will not count toward your presence rate:

• Short or ordinary words. An alias shorter than 3 characters, or one that is also an everyday word, is still matched, but it scores below the 0.5 confidence a mention needs to count.
• Too many terms. The matcher uses up to 200 terms per brand across the name, aliases and domains. Anything past that never produces a mention.

When a save includes either, the dialog stays open and lists the affected terms, so you can edit them or keep them as they are.

​

Archiving a brand

Archiving stops scanning a brand from the next scan. Everything already measured is kept, and past charts keep its name. Archiving cannot be undone, but the name becomes free again, so adding it back starts a new brand. Select Show archived to list archived brands.

​

Page states

When there is nothing to show, the page shows one line and links to the matching section below. Each state is worded separately on purpose: “we have not looked yet” is a different fact from “you are not named”.

​

First scan in progress

A scan is queued or running. Scans trickle across their window rather than firing every call at once, so results arrive over several days, and the numbers stay incomplete until the scan finishes.

​

No scans in this range

Nothing ran in the date range you picked. This is not a measurement of zero. Widen the range, or add prompts on the Prompts tab.

​

Nothing scored yet

Scans ran and answers came back, but no confirmed brand was set up to look for. “Nobody told us what to look for” is not “we looked and found nothing”. Confirm a brand on the Brands tab.

​

No brand of your own

There are answers for this range, but no confirmed brand of your own to measure against. A brand found on your site can be measured, but it is only reported as you once you mark it as yours on the Brands tab.

​

Not measured in this range

Your brand has no scored runs in the selected range, so there is no rate to report. That is different from a rate of zero. It happens, for example, when a brand is added part-way through the range.

​

Scan completed with gaps

Some engine calls did not land, so those windows cover less evidence than a healthy scan. The rates are still correct for what ran. The info mark beside the notice lists the reason recorded for each scan. When engine coverage was removed part-way through a window, the reason says so. The premium engines stopped being called at that point, nothing failed, and the rates are correct for what ran.

​

Shorter evidence range

Raw answers, cited pages and prominence are kept for a shorter time than the rollups behind the rates. When your date range reaches past that, the page names how many days are kept and the shorter range it actually searched, rather than presenting partial evidence as the whole range.

​

Filtered to one brand

With a brand filter applied, runs, coverage and engine counts describe only the rows matching that brand, not the whole site. Clear the brand filter for site-wide totals.

​

Adding prompts and brands

Prompts, brands and competitors are managed on the Prompts and Brands tabs of the AI Visibility page, and through the API and the CLI. Add prompt asks for the question, its type, an optional topic, the locale, and whether it is Active or Paused. A paused prompt costs nothing and keeps its past results. Suggest reads pages from your site and proposes prompts. Nothing is preselected and nothing is tracked until you pick it, and you can pick only as many as your allowance has room for. A prompt’s wording and locale are fixed once created. Stored runs reference the prompt, so editing the text in place would retroactively change what past runs claim to have asked. Edit details changes the question type, the topic and whether it is Active or Paused. Change wording archives the old prompt, keeping its answers, and starts a new one; the dialog says so before you select Create prompt. Your prompt allowance is per account, across every site, not per site. The line above the meter shows both numbers, how many prompts your account uses and how many are on this site, so a site showing two prompts can still be at an account-wide limit. Adding a prompt on any site spends the same allowance.