AI Search Visibility: Choose the Right KPIs
A marketing team can see visits from AI assistants in its analytics account and still have no clear answer to a simpler question: when a potential customer asks an AI tool about the category, does the company appear in the answer? A person might read the brand name without clicking. The tool might cite a company page without naming the company in its recommendation. Or it might name the company while citing somebody else’s comparison. Those are different outcomes, even though a dashboard may label all of them “AI visibility.”

That distinction matters when deciding what to improve. A traffic report can tell you about people who arrived at your site; it cannot count everyone who saw an answer. A list of citations can show where a page was displayed as a source; it does not tell you whether the answer described your product accurately. Semrush recommends reporting AI mentions and citations separately from organic traffic. The practical choice is to use a small set of KPIs that each answer a specific question, with the prompts, platforms, and counting rules visible beside the numbers.
For most teams, I would make brand presence in relevant answers the primary visibility KPI, then read it alongside visible citations, answer accuracy, and visits or conversions. If the commercial question is whether the brand is being considered at all, presence is the closest available measure. If the question is whether the company’s own material is being used as a source, citations take priority. Neither can stand in for business results. The cost of this approach is that the report has several numbers rather than one tidy score; the gain is that each number suggests a different action.
Count an answer before you count a mention
The basic unit of a prompt-based study should be one generated answer to one prompt on one platform at one observation time. A brand can appear three times in an answer without representing three opportunities to be seen. Ahrefs counts a brand mention at most once within a generated response. A team building its own measure can follow that rule: mark each observed answer as brand present or brand absent, then divide the number of present answers by the number of eligible answers. Call the result an answer mention rate, and always show the numerator and denominator.
Suppose, as an illustration, a payroll software company observes 40 chosen prompts five times on one AI platform. It has 200 answer observations. If its brand is named in 60, the mention rate for that platform and prompt set is 60/200, or 30%. That is not a claim that 30% of all AI users saw the brand. The 40 prompts were selected by the company, and the 200 observations are sampled outputs, not a census of user questions. A useful report would say “60 of 200 sampled answers mentioned us” before offering the percentage.
The same answer can contain a brand mention and a citation to the brand’s site, one of those, or neither. A mention answers whether the company entered the conversation. A visible citation answers whether a page was presented as a source. Ahrefs also distinguishes pages retrieved while an answer is prepared from pages visibly cited in the answer; its “found in” measure includes retrieval that did not become a displayed citation. That distinction appears in Ahrefs’ metric definitions. It prevents a common overstatement: a page found in the background should not be described as a source the reader saw.
For a domain citation rate, count an answer once when at least one page from your domain is visibly cited, regardless of how many of your URLs it displays. Then divide by all eligible answers in that same sample. Keep a separate list of the particular pages cited. A high count for one page and a low count for the rest may be more useful for content decisions than the overall domain rate, but the two figures have different denominators. Do not add page citations to brand mentions or call the sum “visibility.”
This answer-level design also makes the awkward cases legible. A brand may be named while a third-party page is cited: the brand is present, but its own site has no visible source placement. A domain may be cited while the company is absent from the recommendation: the page is used, but the brand has not necessarily entered the user’s shortlist. The second case is a reason to read the answer, not to celebrate a citation total. The first case may show that an outside source is shaping the answer, but it does not establish why the model chose that source.
Pick KPIs by the decision they can change
An executive report does not need every count a tool can export. It needs enough separation to tell the team where the next hour or dollar should go. I would start with four measures, then add a fifth when competition is a central concern. Each can be calculated for a defined platform and prompt group.
| KPI | Practical unit | What it helps decide |
|---|---|---|
| Answer mention rate | Answers naming the brand / eligible sampled answers | Whether the brand enters relevant answers at all |
| Visible domain citation rate | Answers visibly citing the domain / the same eligible answers | Whether the company’s pages appear as sources |
| Priority prompt coverage | Chosen prompts with at least one brand mention / chosen prompts | Which customer questions still have no observed presence |
| Description accuracy rate | Reviewed brand mentions with no material error / reviewed brand mentions | Whether more exposure would carry the right information |
| Competitive presence gap | Your mention rate minus a named rival’s rate on the same answers | Where a rival appears more often for the questions you care about |
The first two rates use the same answer denominator, so they can be viewed side by side. They still describe different events. Prompt coverage has a prompt denominator: one appearance in five repeated runs counts as coverage, even if the other four runs omit the brand. Use it to find gaps, not as a substitute for the answer mention rate. The accuracy rate uses only reviewed brand mentions; a team should say how many were read and what qualified as a material error. A competitive gap is meaningful only when both brands are checked against the same answer observations and the same naming rules.
For accuracy, focus on facts that would change a buyer’s decision: what the product does, whom it serves, major limitations, and any claim that your team can check against current company information. “Positive sentiment” is a weaker goal if an answer confidently gives the wrong eligibility or product scope. A simple review code works: correct enough for the user to act, materially wrong, or unclear. Record the passage and the current company source used to judge it. This is a proposed operating measure, not a vendor standard; consistency matters more than a decorative precision score.
If the business sells across several product lines, do not collapse them into a single brand rate without looking at the underlying prompts. A high rate for the well-known product may hide absence from the new product’s questions. Segment by the decisions a customer is making, such as identifying a problem, comparing options, and checking a specific solution. The segments should be chosen because a different result would change the team’s action, not because a dashboard offers an attractive chart.
Build a prompt set you can compare again next month
The hardest denominator is often the prompt set itself. Search queries are uneven: “What is payroll software?” asks for an explanation, while “Which payroll software handles staff in several states?” asks for options, and a question containing your brand name almost invites a mention. Mixing those together without labels can make the overall rate look stronger while the questions that lead to a purchase remain weak. These are illustrative prompts, not reported user queries. Their purpose is to show why the sample must reflect the reader decisions your business wants to understand.
Start with a written set of prompts that represent actual customer questions your team can justify from its own research. Keep branded and unbranded questions apart. Note the intended market and language; a result observed for an English prompt in one market should not silently stand for another. Name the competing brands before sampling, including the spellings and product names that count as an appearance. If a parent company and a product brand are both relevant, decide whether each is measured separately. Otherwise, changes in the alias list can create a rise in mentions without any change in answers.
Keep the observation rule as stable as the prompts: platform, prompt wording, location or language settings available to the observer, whether the run is logged in, observation time, number of repeats, and what qualifies as a visible citation. Those choices define what the reported number describes. Semrush advises fixing the prompt set at the start of a reporting period because adding prompts during the period can inflate counts without showing real improvement. A stable set turns the next period into a comparison. An exploratory set can still be useful for discovering new questions, but its answers belong in a separate view until the team deliberately adds those prompts to a new baseline.
Do not choose only prompts on which the brand already appears. That selection would answer “How often do we appear where we already appear?” A more useful set includes important questions where the company expects to be relevant, including ones on which a competitor appears and the brand does not. The team should be able to explain why each prompt belongs: a customer asks it, it represents a product decision, or it tests a claim the company needs to make clear. Record that reason with the prompt so it is harder to change the sample just to improve the headline number.
There is also a weighting choice. An unweighted mention rate treats every chosen prompt observation equally; it answers how the brand performed across this designed test. If a team assigns more weight to certain questions, it should publish the weights and the reason for them before seeing the results. A weighted rate is still about the chosen test, not a measured share of all AI queries. Without reliable information about how often people ask each question on each platform, a weighted number should not be presented as audience reach. A vendor’s modeled reach may be useful as a directional benchmark, but it should keep its own label and method.
The cleanest comparison is often narrower than the most impressive chart: the same 40 prompts, five observations each, on the same platform in two periods. If a product launch creates a genuinely new customer question, add it to a separate new-prompt cohort and begin its time series then. If a platform changes or a team changes the citation rule, put a visible break in the series. A dashboard that smooths over those changes makes operational decisions worse, even when the line looks easier to present.
Read vendor scores and platform reports on their own terms
Tools can all describe “AI visibility” while counting different objects. Semrush’s AI Visibility metric is a benchmark score distinct from its mention and citation counts. A change in that score may invite investigation, but it is not interchangeable with a percentage of your fixed prompts that named the brand. Ask what population the vendor sampled, which platforms are included, and whether the comparison group changed before treating a score movement as a business result. Keep the raw counts and the score in different rows.
Ahrefs’ mention rule provides a particularly useful reminder about units: it counts whether the brand appears within a response, not the number of times its name is repeated there. Ahrefs’ retrieved-page and visible-citation measures also describe different stages of an answer. Its definitions distinguish both. When moving between tools, check whether “citation” means an answer containing your domain, a cited URL, or every displayed source instance. If the unit changes, two citation totals cannot be laid on one trend line merely because they have the same label.
Platform-owned reporting answers another kind of question. Bing’s AI Performance report lists displayed citations and sampled grounding queries, and says those figures do not indicate page ranking or placement. Grounding queries can suggest the phrases associated with cited content, but they are a sample. A rise in Bing citations can mean that the site is appearing more often as a displayed source within the report’s covered activity. It cannot, by itself, tell you whether your brand was recommended more often or whether the cited page was persuasive to a reader.
Google announced Search Console reports showing impressions and pages for certain generative AI surfaces. A page impression on a Google AI surface has its own value: it indicates that a URL appeared under that report’s rules. It is not the same event as an answer in a separate prompt test naming your brand, and it is not automatically a visit. If Google impressions rise while a third-party prompt tracker shows flat mentions, the first step is to look at the covered surfaces, pages, and samples. There is no sound arithmetic that turns those two measures into a single universal share of AI search.
Google Analytics’ AI Assistant traffic reporting tracks visits associated with recognized assistant referrers. A visit is a later event than a mention or a displayed source: someone has reached your site. A customer who sees an answer and never clicks has no assistant referral visit for that exposure. Conversely, a referred visit does not tell you exactly which phrase or answer persuaded the person. Read this traffic alongside the visibility measures, with its own denominator and conversion data, rather than dividing visits by an unrelated vendor mention count to invent a click-through rate.
The comparison across tools is therefore a map of different stages, not a contest for the one “true” number. A sampled answer test tells you what the selected prompts produced. Vendor benchmarks tell you how a vendor models its covered universe. Bing citations and Google page impressions come from their respective surfaces and rules. Analytics describes visits that reached your property. All are useful when a decision calls for them; none should be silently merged.
Show uncertainty before calling a gain or loss
An answer generated today is not a permanent search ranking. Research on repeated identical prompts found that cited sources and citation rankings can vary across runs. If a company appears once in a single check and disappears on the next, a one-run rank report will overstate both the win and the loss. Repeat the important prompts under the same observation rules and report rates across those observations. Where the sample is thin, show the underlying answer examples as well as the aggregate.
The illustrative payroll test above produced 60 brand mentions in 200 answers, or 30%. Under the simplifying assumption that those 200 yes-or-no observations are independent, a 95% Wilson interval is roughly 24% to 37%. NIST describes the Wilson method for an interval around a proportion. The interval does not make the selected prompts representative of all AI usage. It only communicates sampling uncertainty under a statistical model. Nor should its neat boundaries be trusted if repeated answers to the same prompt move together; those observations may be less independent than the calculation assumes.
For a working report, I would show the count, rate, and range where an interval is appropriate, then inspect how results differ by prompt and by run. If one comparison prompt accounts for most of a change, that is different from broad improvement across the set. If the sample is small enough that plausible ranges for two periods overlap widely, call the movement inconclusive and continue observing. A one-point change with 200 observations is rarely a useful reason to rewrite a content program. A consistent change concentrated in a high-priority prompt group may be worth acting on, especially when the actual answers reveal what has changed.
Uncertainty is not limited to sampling. A prompt set can drift away from the questions customers currently ask; competitors may change names; platform output and citation displays may change. The report should date its snapshot and preserve the prompt list and observation rules. That allows a reader to decide whether the next figure is comparable. If the rules changed, begin a new series or show both methods for an overlap period. Calling a method change “growth” would spend credibility for a number that cannot guide a choice.
Read the answers that drive the numbers
Rates tell a team where to look; the answer tells it what to do. When a brand is present, ask whether the response describes the right product and use case. A sentence that places a company among relevant options is different from a sentence that warns the reader away from it. A mention count treats both as present. A human review of a sample can supply the missing context, provided the report states how many answers were reviewed and which errors counted as material.
Review the source links in the same pass. If the brand is named but only outside pages are cited, those pages may be worth reading to understand how the category is being described. If your page is cited but the answer gives the wrong product detail, compare the answer with the current page before deciding the page needs work. A citation is a clue to the content used in an observed answer, not proof that every sentence came from that URL. The distinction between retrieved and visibly cited pages matters here again: Ahrefs says it can record pages found during generation that were not shown as citations. A team should not present those background pages as displayed endorsements.
The action depends on the failure. Where the company never appears for a clearly relevant comparison question, examine whether its own pages explain the differentiator a buyer is asking about and whether independent sources represent the product accurately. Where the company appears but is described incorrectly, prioritize the incorrect claim over adding more prompts to the dashboard. Where the brand is mentioned and accurately described but the site is not cited, inspect which sources the answer actually shows before commissioning another page. None of those observations guarantees that a specific edit will change an AI answer. They do tell the team which hypothesis is worth testing next.
Avoid treating every missing mention as a content defect. Some prompts may be outside the product’s real scope; an answer that omits the company there may be appropriate. Others may be branded or unusually narrow, making presence easy but commercially unimportant. A monthly review should be allowed to say “we should not chase this prompt.” That judgment protects the report from turning a sampled presence rate into a target that rewards irrelevant appearances.
Connect answer visibility to visits without inventing attribution
AI visibility can matter before someone visits a website, while the business still needs to know whether attention turns into action. The measurement path is a sequence: an answer may name the company, may show a page as a source, may lead a person to visit, and may eventually contribute to a business outcome. Each step loses some people and some information. Semrush separates answer mentions and citations from organic visits in its AI visibility guidance. That separation is a practical guardrail against calling a mention a lead or calling a visit the full measure of exposure.
Use analytics for the outcomes it can actually see. Report AI Assistant visits or sessions, the landing pages those visitors reach, and the conversions your own property records, if those measures are configured and trustworthy. Google’s AI Assistant reporting identifies traffic from recognized assistant referrers and tracks clicks. The denominator for an on-site conversion rate is the relevant visits or users under your analytics definition. It is not the number of sampled AI answers. If referrals rise while conversions do not, investigate the landing experience and the fit of the referred audience before claiming that the visibility work failed.
Branded searches and direct visits may also be useful context, but they are not a clean receipt for an AI mention. A campaign, news event, product change, or other channel can move those figures at the same time. Track them alongside the answer data and label any relationship as an observation, not a causal return on investment. If leadership asks for revenue attributed to AI search, the honest answer may be narrower: revenue from identifiable assistant referrals can be reported under the company’s existing analytics rules, while the value of answers seen without a click remains unobserved. That limit should change the wording of the claim, not stop the team from measuring what it can.
Make the monthly report point to the next decision
A useful monthly page could begin with a fixed set of 40 priority prompts on one platform, observed five times each: 200 eligible answers. Suppose the brand appears in 60 answers and its domain is visibly cited in 22. The report would show a 30% answer mention rate and an 11% visible domain citation rate, with the counts beside them. If reviewers read 30 of the 60 brand mentions and find four with material product errors, they should write “4 of 30 reviewed mentions were materially wrong,” not “4 of 200 answers were wrong.” These are illustrative numbers; their value is in keeping each denominator attached to the claim.
The next line should explain where the change occurred. Did the brand enter more unbranded comparison answers, or did a branded question account for most of the gain? Did citations come from a page that answers the customer’s question, or from a page whose information is out of date? Did a named competitor’s presence change on the very same answer sample? The report can show the prompt groups and a few representative answers so that the headline rate leads to a choice rather than a congratulatory graph.
I would then choose one response to the strongest pattern. If relevant answers rarely name the brand, prioritize the product information and outside descriptions that address those prompts. If the brand appears often but material errors recur, fix the information that people and AI systems can read about the product, and check subsequent answers for that exact error. If visible citations rise while referrals stay flat, keep the citation and traffic results separate and inspect the cited pages and calls to action. If only a vendor score moves while the fixed prompt results do not, investigate the vendor’s scope before reallocating budget. Each choice has an opportunity cost, so the pattern should be specific enough to justify it.
The strongest KPI set is small because its job is to preserve meaning. For a defined prompt set, report whether the brand is named, whether its domain is visibly cited, whether the description is sound, and what visits and outcomes your site actually records. Keep platform reports and vendor benchmarks in their own units. Repeat observations before treating a movement as a trend. That gives a team a defensible answer to “Are we visible in AI search?” and, more usefully, tells it what to improve next.