How Trovance measures.

Five live engines, each asked through its developer API, not the app a person signs in to. Google AI Overviews are the exception: Trovance doesn’t ask them a question, it reads them off a Google results page.

• The OpenAI API, with the model gpt-4o-mini and OpenAI’s web-search tool offered, not forced. This is not the ChatGPT app.
• The Perplexity API, with the model sonar.
• The Gemini API, with gemini-flash-latest, a rolling alias, and Google Search grounding.
• The Claude API, with claude-sonnet-5 and web search, up to two searches an answer.
• Google AI Overviews, read from a Google results page for the United States, in English, on desktop, through the search-results provider DataForSEO.

Beside them, a Claude memory check asks the same Claude model, with search turned off, to answer from its own knowledge. It’s kept out of the headline numbers.

Your plan sets how many live engines run. Starter covers the OpenAI API and Perplexity; Growth and Agency cover all five. Every plan includes the Claude memory check. A trial runs Starter’s engines until a teammate is invited, which moves it to Growth’s for the rest of its 14 days.

Each question goes to each engine once a week. Weeks run Monday to Sunday in UTC, and the scheduled sweep starts on Monday at 02:07 UTC. A manual re-run in the same week can’t add a second answer: Trovance keeps one per question, engine and week.

Claude with search is the one engine that doesn’t answer every question every week. In most weeks it answers a fixed quarter of them, the same quarter each time, and in the first sweep of each month it answers them all. The weekly note says so: “A smaller mention count here is the sample, not a market move.”

A language model writes them from what’s saved about your brand: its offerings, audiences, markets and topics. It writes category, alternatives, how-to, problem and comparison questions, and it’s told to leave your brand’s name out unless a comparison needs it. Up to 10 of your Search Console queries can seed more, and your team can pin questions of its own.

They’re hypotheses Trovance asks on your behalf, not questions buyers were seen asking. In the app’s words, generated questions “are hypotheses, not measured market demand.”

Your plan caps the count: 25 questions on Starter, 100 on Growth and Agency. About 80% of the cap forms a stable core, with pinned questions first. The rest are exploration questions, which rotate from week to week; questions seeded from Search Console are always exploration. Only core answers count toward the headline numbers.

Each week, up to five questions per brand get repeats, the ones whose recent answers have been mixed first. Each gets two more samples on every engine that answered it, three in all. In the app’s words, three samples “describe variability; they do not establish a precise probability,” and repeats never change the headline counts.

About half of your questions are a control: they’re asked like the rest, but Trovance never files a recommendation or a draft from them, and the weekly note gives the split. It doesn’t yet turn that control into an estimate of noise.

For each answer, Trovance keeps the text, up to its first 8,000 characters; the sources the engine attached as citations, kept apart from the pages it retrieved; which tracked brands were named or cited; and a record of the exact question, model and request settings. Named and cited are judged on the whole answer, before it’s trimmed.

A brand is named when its name, its domain, or its domain without an ending like .com or .ai (when what’s left is four letters or more) appears in the answer as a whole word. It’s cited when a source the engine attached links to its domain or a subdomain. These are two separate judgments, made by exact matching, not by a model.

Naming is presence only. The weekly measurement records no sentiment, no recommendation and no position: an answer that names you while recommending a rival counts the same as one that recommends you.

Share of answer is named answers divided by answers received, on core questions, with the live engines pooled by adding up their answers rather than averaging percentages, rounded to a whole percent. Claude memory is left out. Supported, shown beside it, is the same count for cited answers, drawn only for weeks with complete citation data.

When Google shows no AI Overview for a question, that counts as an answer that didn’t name you: in the app’s words, a topic Google declines to summarise “pushes this figure down exactly as a rival winning the answer would.” Any core questions that name your brand count too; the app’s Answer Ledger keeps those apart from questions that never name you.

If an engine call fails, times out or is refused, nothing is stored. That question is absent from the week’s count rather than counted as not named. After three failures in a row Trovance stops asking that engine for the sweep, and the weekly note says its “numbers are absent, not zero.”

A change from the week before is stated only when both weeks have at least 10 answers, the number of answers moved by no more than 25%, and the questions, models and search settings match. Otherwise the app shows the reason instead of a number.

Rivals are matched the same way, on every answer. Your rank among the rivals you track counts the answers that name or cite each one, and ties go against you.

The free check is a first look, and narrower than the platform. It drafts three questions from your site, written to leave out company and product names, and you can edit them before an engine is asked. Then it asks one engine, normally Perplexity, each question once.

It marks each answer “you appear” or “absent,” using a simpler test than the weekly sweep: your domain in the answer or in a source link counts, and so does your name in the answer, unless the answer says something like “could not find” or “not familiar.” So “you appear” isn’t the same measure as named. It records no citations.

If you’re missing from at least one answer and a competitor’s web address shows up in an answer’s short excerpt, the summary names the one that shows up most. The three questions are never asked again, so they can’t be compared with a later week.

The app’s own line: “One engine and three questions. This is an initial signal, not a complete benchmark.”

One thing. For an article drafted and published through Trovance, with its live address recorded on your own domain, Trovance fetches that address and finds the complete saved article text, in order, in the server’s response. The page must load and carry most of the article’s title too. That match is recorded as verified.

The fetch is a plain HTTP request with Trovance’s own user agent, TrovanceVerifier/1.0. It runs no JavaScript, so it reads what the server sends, not what a browser renders. The address must be on your domain or a subdomain, and every redirect is checked against the same rule. In the app’s words: “The complete saved article text was found in order in the server response.”

The check runs when someone verifies the article in the editor or an agent reports its live address, and again in Trovance’s weekly analysis cycle for articles it’s following up on. Publishing alone never counts as verified, and a failed check records the attempt and leaves the article where it was.

Verified doesn’t show that an AI engine has read, indexed or cited the page.

For an article written from one of Trovance’s recommendations to counter a rival, claim an answer nobody holds, or refresh a page, Trovance compares a window before publication with a window after verification, at 7, 28 and 56 complete days, for the exact published address. The days between publication and verification are left out. It compares:

• search clicks, impressions and click-through rate, from Search Console;
• page views and engagement time per view, from GA4;
• AI mention rate and AI citation rate, at 28 and 56 days only, over a fixed weekly panel: the article’s target question, on each engine that answered it with search in the last full week before publication, leaving out AI Overviews and Claude memory.

The search and GA4 comparisons need those accounts connected. A comparison is withheld unless both windows are complete, come from the same source and timezone, and clear minimum samples, such as 20 clicks or 50 page views. Missing days are never counted as zero.

A result can be positive, negative, mixed or inconclusive. A metric that moved carries the sentence “This is a before-and-after association, not proof that the intervention caused it.” Every stored result also lists its limits, starting with “Before-and-after association only; other campaigns, market changes, seasonality and tracking changes are not controlled.” Conversion rate and revenue aren’t compared before and after at all.

A 28- or 56-day result can become a lesson that moves a later recommendation’s priority by at most 3 points, for the same page or the same question only, and never for a request a person made. Lessons expire, a later checkpoint replaces an earlier one, and a teammate can withdraw one with a reason.

It starts with the visit. A visit is an AI referral when its landing address carries a recognized AI utm_source tag, or it arrives from a recognized AI assistant’s site; the tag is checked first. The class is decided when the pageview arrives, and stored.

Trovance’s pixel sends pageviews and heartbeats, nothing else, with no cookies, local storage or fingerprinting. Leads and sales come from your own server, through Trovance’s conversion API, or from Stripe. The pixel’s visitor identifier includes the UTC day, so it counts visits within a day, not people over time.

A sale is traced to an AI-referred visit when a server-side join finds one of four links:

• an AI-referral pageview by the same daily visitor hash, on the same UTC day, no later than the sale;
• for a sale with no such visit of its own, the visit the customer’s first recorded lead was traced to, when the sale carries the same customer reference and comes within 365 days of that lead;
• a Stripe checkout’s client_reference_id, when it’s the pixel’s event id for one of your pages;
• for a renewal, the visit the customer’s first recorded Stripe checkout was traced to, within 365 days of that checkout or of the lead it was traced through.

The UTC day is a hard boundary. Two people can share a daily identifier, so a sale can be traced to someone else’s visit; the receipt then says it was joined through the same visitor that day, never as the landing itself.

Revenue counts one source per brand: Stripe while a Stripe webhook is connected with its currency saved, otherwise your own server. The other source’s revenue is shown beside it, never added. Stripe revenue is live mode only, paid checkouts and renewal invoices only, and USD only. Refunds are not subtracted.

“Traced to” isn’t a claim about cause. The app’s help says so: “A conversion is traced to an AI-referred visit; that is not a claim the visit caused it.” Its receipts say “traced to”, “traced through” or “traced by”, never that a visit caused a sale.

• The weekly measurement doesn’t record where you sit in an answer: not your position, your sentiment, or whether the answer recommended you. It records named and cited.
• It doesn’t see what every user, model or region sees. No persona, location or account goes out with a question (a question can name an audience or a place in its words), and only AI Overviews has a locale (United States, English, desktop). Sampled API answers can differ from what a signed-in person sees in a consumer app, and with no temperature or seed set, from one run to the next.
• It can’t credit Gemini’s citations to a site: Gemini cites through Google redirect links that hide the publisher.
• It doesn’t track AI crawler visits yet.
• It doesn’t see all AI traffic. Referral counts are a floor: an AI browser that looks like ordinary Chrome can’t be told apart, and clicks from Google’s AI Overviews arrive as Google search.
• It doesn’t compare conversion rate or revenue before and after a change.
• These numbers don’t score your site. Agentic Web Health does that separately, with its own weighted checks: the August 15 release note shows the rubric, and the app sets out how each check is scored once you’re signed in.