Skip to content
Method

How we measure mention costs.

This page is why the cost of a mention is built the way it is, what in them is measured, how the one quantity that could not be read off a log was measured, and what the figures still cannot tell you. It carries no figures of its own, deliberately: they belong next to the companies they describe.

← Back to Results

What the figure is made of.

Four inputs, three of them measured directly.

C = P / M, where M = (L / k + F × R) × m

  • L, live AI reads. An assistant retrieving a company’s content while it composes an answer. Measured, from our own request logs, identified by the requesting agent.
  • k, pages read per live answer. Measured, from the same logs. A request is not an answer: an assistant that reads two pages while composing one answer logs two requests, so dividing by k is what makes the first term a count of answers rather than of fetches. Treating requests as answers, which is the form this estimator usually takes, overstates answers and understates cost.
  • R, index refreshes. An AI search crawler re-reading that content to store it for later answers. Measured. Training crawlers are a separate population and are excluded entirely: they feed model weights, not an answer index.
  • F, answers served per index refresh. The one quantity that cannot be read off a log at all, because it happens inside an assistant. How it was measured instead is the subject of the rest of this page.
  • m, mention rate. The share of answers in which the brand appears, measured on that company’s own monthly question set and taken from its own report.
  • P, the monthly price that company pays. Real, never a reference figure, and never published.

The paid click it is set against

  • WordStream by LocaliQ, Google Ads Benchmarks 2026: median cost per click across 13,474 United States search campaigns in Google and Microsoft Ads, 23 industries, April 2025 to March 2026, in US dollars. The figures are medians of that publisher’s own managed client campaigns, not a market average: across its 23 industries they run from 1.63 to 9.87 dollars.
  • Dreamdata, LinkedIn Ads Benchmarks Report 2026: cost per click 5.98 euro, measured over 66 million sessions in 3.5 million completed B2B customer journeys, August 2024 to July 2025. The same report puts Meta at 1.60 euro per click, the cheapest paid click in it.
  • Each company is matched to the category its own business sits in, so its paid-click comparison is read against that category’s figure rather than an average across all of them. The dollar benchmarks are converted once, at the European Central Bank euro reference rate, 16 Oct 2026, so that the two figures and the difference between them are in one currency and divide as they read.
  • A published click price is what advertisers paid for a click in a competitive auction. It is not what a click is worth, and neither figure in that comparison is a return.

Neither P nor M is published, and neither is any request or refresh count. What is published is the quotient C. Stated precisely rather than as a guarantee: a reader who also knew a company’s price could solve for its estimated answer volume, which is why the price is the thing these pages never give and never point at.

The one quantity that had to be measured.

F cannot be counted from outside an assistant. It is also not a matter of opinion.

Nobody outside an assistant vendor can count the answers a stored page goes on to serve: we observe our own server, not their output. F is nonetheless tied to a quantity that can be measured, and the tie is an identity rather than a model.

F = (L / (k × R)) × (1 / f − 1)

  • f is the share of answers that use a company’s content and read it live at the moment of answering, rather than answering from what was stored earlier. The derivation is immediate: a live answer makes k reads, so live answers are L divided by k; if a fraction f of content-using answers read live, total content-using answers are L divided by k times f, and the rest were served from refreshes.
  • L, R and k are all measured, so F is a function of exactly one unobserved quantity. Any claim about F is a claim about f, and a measurement of f settles F. That reduced the whole question to measuring one share, which is what the work described below did.
  • Substituting the identity back into the estimator removes F altogether: L over k plus F times R is simply L divided by k times f. The published cost is therefore proportional to f, one for one. A relative error in f passes through as the same relative error in the cost, with no amplification at either end. The parameter that looks fragile is not the one the published figure depends on.
  • The upper end is fixed without any measurement at all. At f equal to one, every content-using answer reads live, nothing is served from storage, and the cost is the no-reuse figure. That is the ceiling, stated here rather than left unsaid, because it is the number that is true if index reuse turns out to deliver nothing.

How f was measured.

Three stages, designed and written down in full before the first observation, so the design could not be fitted to the result. The third is randomized, which is what moves the central claim from estimated to established.

  • Stage one. Does a refresh serve an answer at all?

    A pre-registered natural experiment that published nothing new. From our own request logs we identified layer pages that no AI index crawler had ever fetched, together with questions those pages were uniquely able to answer, and verified that the answer was not findable anywhere else. Those questions were put to the live assistant surfaces first, establishing a baseline that had to come back empty. The logs were then watched for the first index-crawler fetch of each page, recorded to the second, and the same questions were asked again afterwards in varied phrasings from fresh sessions.

    What it reads out. An answer that returns the page’s content after that first crawl, with no live read of the page in our logs at the moment the answer was produced, is a direct observation of an index refresh serving an answer. Run per assistant, it answers the question a pooled parameter has to assume away: whether every assistant’s refreshes feed answers, or only some assistants’ do. An assistant whose refreshes never produce one leaves the estimator entirely rather than being down-weighted.

  • Stage two. How many, and for which assistant?

    A controlled test on an isolated subject rather than on a customer. The subject is one we created and published ourselves and that did not exist before we made it: it has no website of its own, no press, no listings, no competitors working on the same ground, and nothing about it in any assistant’s earlier training. Its baseline is therefore zero by construction rather than by measurement, and an answer containing it can have come from one place only. What is new is the subject, not the ground it stands on: it is published into a topic that is already asked about and already well answered by others, so the test is contested even though the attribution is clean. It is a real subject and not a fiction. We publish something that genuinely exists because we made it, described as exactly what it is, because inventing a company or a product and getting assistants to repeat it would put fiction into AI answers, which is the thing this entire measurement exists to avoid doing to anybody. The questioning ran to a number of qualifying answers fixed in advance rather than to a deadline, so the precision was set before the result was known rather than chosen after it.

    What it reads out. It measures the share of content-using answers that read the page live at the moment of answering, free of every attribution problem the rest of this page has to argue around. That share is observable because all three ends belong to us: the subject, the content and the request log. The parameter follows from it directly, computed per assistant and weighted by each assistant’s share of refreshes. The subject is new; the questions it answers are not. They are chosen because they are already asked and already have well established sources answering them, and both are verified before the subject exists rather than claimed afterwards. So the test runs in a contested, demanded topic and still has perfect attribution, which is the combination a test on a real company cannot give. The questions are fixed in advance for demand and for the strength of what already answers them, not for how winnable they look, because a topic chosen after the fact for being easy would measure the choosing rather than the layer. What remains out of reach is worth to a brand: a new artifact has no brand equity, nobody searching for it by name and no purchase intent attached to it, so it cannot say what a layer is worth to an established company. That is measured on real customers against the comparison group above.

  • Stage three. What does the layer add over simply having a website?

    A randomized trial, run on subjects of our own rather than on customers, so that nothing anybody paid for is held back or altered. Several real subjects are created, each one a genuine artifact that did not exist before, and every one of them is given an ordinary web presence of the kind any company already has. A layer is then built for a randomly chosen half, with the same machinery that builds a customer’s, copied rather than modified. The other half keeps the website and nothing more. Which subjects fall into which group is decided by chance before anything is published, and the questions, the sample size and the stopping rule are fixed at the same time.

    What it reads out. The difference between the two groups is the effect of the layer over and above having a website, which is the comparison a buyer is actually making. Because assignment is random, that difference is causal: the two groups cannot differ systematically in anything else, since nothing decided which subject went where except chance. This is the one thing no measurement on customers can deliver, because a company chooses to become a customer and that choice can never be randomized away. What it does not settle is transfer: these subjects are new, so they carry no brand equity, no established rivals of their own and no history, and the size of the effect on an established brand need not match the size measured here. That size remains an estimate, and a published figure is labeled as one.

  • The rule was fixed first. Whatever share the measurement produced, the published figures take the measured value, and they do so whichever direction it moves them. The acceptance rule, the stopping rule and the number of qualifying answers were written down before the first observation rather than chosen once the shape of the result was visible.
  • It could have raised our own costs, and it still can. A smaller share of live reads means more answers served from storage and a lower cost per mention; a larger one means the opposite. The design was recorded in advance precisely so that an outcome which makes our figures look worse cannot be quietly dropped, and the same rule applies to every repeat.
  • It is repeated, not done once. Assistants change how they retrieve, and a share measured once would decay into an assumption again. The measurement is re-run, per assistant, and a published figure carries the window it was measured in.
  • What is deliberately not described here is the operational detail: which pages, which facts, which phrasings, and how an answer is matched to a request. That is withheld because publishing it would let the measurement be both copied and defeated, not because it is unspecified. It is recorded in full internally, and the record is dated.

What counts as a mention the layer added.

A cost per mention and a cost per added mention are different numbers. The figure published for a company is the first of the two. This section is what the second is measured against.

The mention rate counts every answer that names the brand, from any source. Dividing a fee by that rate prices the brand’s present position in AI answers. It does not price what the layer contributed, and for a brand that was already well known those two are far apart. Two questions follow immediately, and this section is the answer to both: what the layer is credited with, and against what.

  • The baseline is a real pre-period, not a reconstruction. Every customer is measured once before its layer can have had any effect: the first run is dated to the month the domain was registered with us, and the layer’s first AI request comes only after it. That first measurement is the brand’s own level without the layer, taken with the layer’s own instrument.
  • The questions do not change between the two points. The baseline and every later run put the same fixed set of questions to the same assistants in the same phrasings. A before and after measured on different questions would not be a comparison at all. What cannot be held fixed the same way is the assistant behind them: only the changes a vendor labels are visible to us, so a retrieval change made behind an unchanged model name would pass unnoticed here, and only the comparison group described below would absorb it.
  • Where the instrument did change, that arm is separated rather than averaged in. Assistant vendors retire and replace the models behind their own products, and when one of the measured models is replaced, a rise in the rate on that arm can be the new model rather than the layer. The added figure is therefore computed on the arms whose model is unchanged across the two points, and an arm whose model was replaced is reported separately and labeled. The arm is held out whichever way it moves, and that cuts both ways rather than ours: a rise there may be the replacement rather than the layer, and a fall there would hide a rise the layer did cause. Holding it out means forgoing movement we might otherwise have claimed.
  • A before and after is still not a controlled experiment, and one customer cannot run the same month twice. Anything else that moved in the period, the company’s own marketing, its press coverage, a competitor leaving the field, sits inside the difference. That is why a control arm exists at all: a standing panel of brands that are not customers, in the same categories, measured on comparable question sets on the same schedule. The movement in that panel is what a customer’s movement is read against, so market-wide drift is subtracted rather than claimed.
  • What the control arm cannot do is make the comparison an assignment. Customers chose to become customers, and a brand that buys AI visibility is not a random draw from its category. The panel removes drift common to the category; it does not remove the difference between a company that acted and one that did not. The added figure is therefore reported as an upper estimate of what the layer contributed, never as a causal claim, and the cost per added mention is never lower than the cost per mention it is read against. A company whose measured rate has not risen above its baseline has no added figure at all.

The two are never presented as one. The cost per mention is what the brand’s position in AI answers costs; the cost per added mention is what the movement since the baseline costs.

What is established, and what is estimated

  • Established by construction: that a layer causes AI assistants to carry its content into answers, how long that takes, and how often they read it live rather than from storage. The isolated test settles these, because its subject exists only because we published it and nothing else in the world refers to it. There is no attribution question there to answer.
  • Established by random assignment: that a layer adds mentions over and above simply having a website, and by how much on subjects of that kind. The randomized stage settles this, because chance alone decided which subjects received a layer. That is a causal effect with a control group, and it is the claim a buyer is really asking about.
  • Estimated, with a named limit: that the effect measured on test subjects is the size an established brand would see, and that a particular customer’s rise in mention rate was owed to its layer. Neither is established. Test subjects carry no brand equity and no history of their own; and a company chose to become a customer, which no comparison group turns into an assignment. The control arm removes what moved across the category, not the difference between a company that acted and one that did not.
  • Why they are stated apart. Read together they collapse into the weakest one, and a mechanism that has been isolated and a trial that has been randomized end up sounding as unsettled as the hardest claim on the page. They are claims about different things. What a layer does, and what it adds over a website, are settled where they can be. What a particular brand’s rise is owed to is not, and the added figure is published as an upper bound on it rather than as a measurement of it.

The objections, and what each is worth.

Every serious objection we know of, hardest first, with the concession before the argument. Several of them cannot be closed, only bounded, and those say so in those words rather than leaving the reader to notice.

  1. A cost per mention is not a cost per mention you added. It divides the fee by mentions the brand already had.

    Conceded. Correct as an objection to a cost per mention, which is why it is not the only figure measured. The mention rate counts answers naming the brand from any source: the company’s own website, retailers, press, review sites, an assistant’s own training. Divided into a fee, it prices the brand’s present position, not the layer’s contribution to it.

    The argument. So the added figure is measured separately, against the brand’s own pre-layer baseline, on the same instrument, with the arm whose model changed held out, and read against a panel of non-customer brands in the same categories. The section above sets out how. What remains is the limit of any before and after: customers chose to become customers, so the comparison removes category drift but not self-selection, and the added figure is published as an upper estimate of the layer’s contribution rather than as a causal claim. It is never the lower of the two costs, and the two are never reported as one number.

  2. If you want to know what the layer caused, measure only the customers who run nothing else alongside it. Then the increase is yours.

    Conceded. That is the right instinct for the added figure, and it was the first thing we tried. It fails on two practical grounds before it reaches the hard one. Whether a company is also running a press push, a product launch or an agency retainer in the same quarter is something we would be asserting rather than something a reader could check. And the companies most likely to qualify are the smallest and least active ones, which would make the measured effect a property of that sample rather than of the layer.

    The argument. Underneath both sits the real obstacle: there is no untreated group in our data to compare against. A company is only measured once it is a customer, so every brand in the set is already being worked on, and none of them is observed over time without a layer. Comparing early adopters against later ones does not rescue it either, because the later ones are not measured before they arrive. So isolation is used, three times over, and in none of them by picking customers. For the mechanism it is taken further than selecting quiet companies could ever take it: the second measurement above runs on a subject with no company behind it at all, created by us, with a baseline of zero by construction and no other party able to influence it. For what the layer is worth to an established brand, which a subject with no brand equity cannot tell you whatever topic it sits in, the comparison group is built from outside the customer base: brands in the same categories that are not customers, measured on the same schedule with comparable question sets. That is an actual control arm rather than an assertion about what a customer does in private. And for the question underneath all of it, what a layer adds over simply having a website, the third stage above does what selecting quiet customers was only ever an approximation of: it assigns by chance. Nothing a customer receives is altered to make that possible, because the product is replicated on subjects of our own rather than withheld from anybody who paid for it.

  3. Your own measuring models were replaced during the period. A before and after measured across that is worthless.

    Conceded. For the arm where it happened, that objection is correct, and it is the reason the arms are reported separately rather than pooled. An assistant vendor replacing the model behind its own product can move a measured rate by itself, with nothing about the brand having changed at all.

    The argument. So the added figure is computed only on the arms whose model is identical at both points, and the arm whose model was replaced is reported separately with the replacement named rather than quietly averaged in. The rule is not the convenient one: pooling the arms would hand the layer credit for whatever a replacement moved on its own, and holding the arm out forgoes that credit rather than protecting it. The arms are decided at computation time from what each individual answer was actually produced by, not from a list someone maintains, so the next replacement is handled the same way without anyone having to remember.

  4. Does that attribution problem not apply to the answers-per-refresh figure as well?

    Conceded. It is a fair question, because the two measurements sit next to each other and look alike, and because this page spends most of its length on attribution.

    The argument. It does not apply, because that figure is not about the customer. How many answers one index refresh goes on to serve is a property of the assistant: how often it answers from what it stored rather than reading again. It does not move because a company also runs search ads, changes agency or launches a product, and there is nothing in it to attribute to anyone. That is why it is measured on assistant behavior, with content we control and a request log we own, and why isolating customers would neither be required for it nor improve it. The attribution question belongs entirely to the added figure above, and it is answered there with a comparison group rather than with an assumption.

  5. A mention is not a click, so comparing their costs is not like for like.

    Conceded. Correct, and we do not claim otherwise. A paid click delivers a visit; a mention delivers an appearance inside someone else’s answer. Pew Research measured a click on a link inside a Google AI summary on about 1% of visits to a page carrying one, so a mention is emphatically not a click bought at a discount, and nothing here treats it as one.

    The argument. The comparison is nonetheless the relevant one, because the click is leaving the place where buyers now decide, and four published measurements say so rather than us. Forrester, surveying roughly 18,000 business buyers worldwide for The State Of Business Buying 2026, found 94% had used AI somewhere in their buying process, up from 89 percent the year before, and twice as many named generative AI or conversational search as a more meaningful source than named any other. SparkToro, on Similarweb clickstream data for the United States from January to April 2026, found 68% of Google searches ended without a click of any kind, up from 60 percent two years earlier. Ahrefs, across 300,000 keywords of aggregated Search Console data, measured 58% fewer clicks for the top organic result on queries that show an AI Overview in December 2025, against 34.5% a year earlier. Similarweb found that 6.8% of ChatGPT answers carried even one external link in May 2026, up from about 1.3 percent a year before, which means that in more than nine answers out of ten there is no click to win at all and the only question is whether the brand is named. Where the click does still happen it arrives later and warmer: three independent multi-site measurements put AI-referred conversion at 1.18 to 1.26 times organic search, and Adobe measured AI-referred retail traffic converting 60 percent better than non-AI traffic in July 2026, the eleventh consecutive month it came out ahead. That last figure moves about, and the same panel had it 38 percent worse in March 2025, which is why the claim made here is the modest one: a mention is not a click, but it is the unit that still exists on the surface where the decision is now being made, and what does arrive from there converts at least as well as search.

  6. The number of answers an index refresh goes on to serve is the whole figure, and you cannot see inside an assistant to count it.

    Conceded. The second half is true and it is the reason this page exists. It is the estimator’s single most consequential term, every cost figure is inversely proportional to it, and no amount of reading our own logs can observe it directly.

    The argument. So it was not read off a log. It was reduced to one observable share by the identity above and then that share was measured, per assistant, by the two-stage design below. Before that measurement existed the term was bracketed four independent ways, each on different data and different external literature: a revealed-click triangulation from logged AI referrals against published per-citation click-through rates; a crawl-economics derivation from re-crawl cadence on a maturity-matched cohort; a top-down demand derivation from published category population denominators and published AI adoption and retrieval rates; and a transfer of published crawl-to-referral ratios corrected for the search-index share of AI crawling, with two further adversarial re-derivations attacking all four. Those brackets are not what the figures rest on now; they are what made it possible to state in advance which outcomes would be surprising. What is published is the measured value, the window it was measured in, and the no-reuse ceiling.

  7. Those four derivations are not independent if they all read your own logs.

    Conceded. Three of them do, and they therefore inherit the same decision about which requesting agents count as an index refresh rather than as a training crawler. A misclassification there would move three of the four together. Only the top-down demand derivation is external end to end.

    The argument. Which is why the classification is not ours to assert: it follows each vendor’s own published declaration of what each of its agents does, and training crawlers are excluded from the estimator as a separate population rather than discounted inside it. More to the point, those four brackets are no longer what the figures rest on. The measurement below observes an answer appearing after a refresh regardless of how the refresh was labeled, so it is independent of the classification decision in exactly the direction that mattered.

  8. The mention rate is measured on a fixed question set and then applied to live AI requests, which are not the same population.

    Conceded. This transfer is not validated. It is the second-largest assumption in the estimator, and the rate it transfers is a company’s own measured mention rate.

    The argument. It is bounded in three ways, and none of them makes it go away. It is linear: an error of a given percentage in the rate moves every cost figure by the same percentage, with no amplification, unlike the parameter above it. Its direction is knowable rather than open: the monitored questions are unbranded buyer questions asked without naming the company, so the measured rate is how often the brand appears unprompted, which is a lower bound on how often it is named in a conversation a buyer has already steered toward it, and a lower bound on the rate is an upper bound on the cost. And it is the assumption that the three stages above do not touch, which is why it is still listed here as an open one rather than quietly retired along with the others. What would settle it is a measurement of the rate on real answers rather than on asked ones, which requires seeing the answers themselves and not only the request that fetched the page. We cannot see those, and we say so instead of modeling our way around it.

  9. Publishing more pages lowers the cost per mention without improving anything for the customer.

    Conceded. True, and it is a real defect of the metric rather than a misreading of it. More published pages attract more index refreshes, more refreshes raise estimated answers, and the cost per mention falls while nothing about the customer’s position in an answer has changed.

    The argument. Two answers, one of which is new. The first is to refuse the use the defect enables: this figure states what a month that already happened cost, per company. It is not a performance indicator, it is not reported to customers as a target and it is not optimized against. The quantity that does measure the customer’s position is the mention rate, measured on a question set of fixed size, which does not move when more pages are published, and a company’s place in the cost order follows its request volume as much as its measured visibility, which is why the mention rate is always published alongside the cost figure rather than folded into it. The second is that the objection contains an empirical question rather than only a complaint, and the randomized trial answers it. Treated subjects there are assigned different amounts of published content, at random, so the trial measures whether more content actually produces more mentions or only more estimated answers. If it does produce more mentions, then a falling cost per mention is the thing working rather than the metric being gamed. If it does not, the metric is inflatable and we change it. That is a better position than defending it, and it is the reason the question was built into the trial rather than answered in prose.

  10. The companies named are hand-picked, so the figures are not representative.

    Conceded. The companies named are indeed a selection, and we will not pretend they are a sample. An earlier approach made it worse by selecting on the outcome, naming only the companies with the lowest cost per mention; it does not do that any more. What remains is that naming a customer requires that customer’s permission, so the list of names is shorter than the list of companies behind the figures and always will be.

    The argument. Two separate things were being confused, and they are now separated in the arithmetic as well as in the words. The first is which companies the estimator prices, and that has nothing to do with how the figures come out: every customer it can price is priced whose own latest report measured at least a stated minimum number of recommendations, a precision test rather than a price test, applied before any cost is looked at. A company sits outside that set only because the estimator cannot price it, for one of four reasons fixed in advance: no AI requests inside the window, so there is no denominator; a latest report older than the window, so there is no rate measured inside it; a measured rate of zero, which prices no mention; or a contract in another currency, because converting one would add an exchange assumption to a figure that has enough of them already. The second is which companies are NAMED, and that is a permission each company gives or withholds. It is a selection, openly, and it would be absurd to pretend otherwise. What matters is that the two never touch: every published statistic, the median included, is taken over the whole priced set before anyone is named or left out, so the headline cannot move because a company agreed or declined to appear. The build refuses a page on which those two sets would produce different statistics.

  11. AI answers are not reproducible, so a rate measured from them is noise.

    Conceded. A single answer is close to worthless as evidence. Published repeat-run studies find that the set of sources an assistant cites for one prompt only partially overlaps with itself between runs.

    The argument. Which is why nothing here rests on a single answer. The mention rate is the share of a fixed, repeated set of answers per monthly run, across several assistants and several phrasings of each question, and the same set is asked again every month. Averaging over a set of that size is the standard treatment for exactly this variance, and it is why the rate is reported monthly rather than continuously.

  12. Assistants differ enormously in how often they read a page live rather than answer from an index, so one pooled parameter cannot be right for all of them.

    Conceded. The objection is correct about the premise: measured across assistants, the ratio of live reads to index refreshes spans several orders of magnitude, and a single pooled parameter would be a simplification that the data itself contradicts.

    The argument. Which is why the parameter is not pooled. The share is measured per assistant and weighted by each assistant’s share of index refreshes, so an assistant that refreshes heavily but rarely reads live contributes according to what it was observed to do rather than to an average of everyone. An assistant whose refreshes were never observed serving an answer is removed from the estimator rather than down-weighted. What remains unmeasurable is the tail: an assistant we see too little of cannot be characterized, and its refreshes are carried at the fleet value with that stated rather than silently.

  13. Traffic from AI assistants is a rounding error against the web as a whole, so this prices something negligible.

    Conceded. Measured as referral traffic it is, and independent measurements agree on that.

    The argument. Referral traffic is the wrong denominator, and it is the denominator this page argues is disappearing. What is being priced is the appearance in the answer, which happens whether or not a referral follows, and which by the published figures happens in the large majority of answers that carry no link at all. The relevant measure of size is how many people are asking, not how many click through, and on that measure the published user numbers are large and still compounding.

What the figure is not.

The list is short and it is exhaustive.

  • Not a cost per visit, per person, per conversation or per customer. Requests are not people.
  • Not a measured count of real-world mentions. It is built from request data and a measured rate.
  • The cost per mention is not a price for the mentions FAIND added: it divides the fee by every answer naming the brand, including answers it would have been named in anyway. The cost per added mention, measured against the brand’s own pre-layer baseline, is a separate figure and is never lower.
  • Not an upper limit. The upper limit is the no-reuse figure.
  • Not a measured business result, and not a claim about revenue, leads or conversion.
  • Not a ranking of AI visibility. A company’s place in the order follows its request volume as much as its measured mention rate, which is why the mention rate is published in its own right, not folded into the ranking.
← Back to Results

Get your free AI Visibility Report

Enter your work email. We'll run a thorough analysis of how AI sees you across ChatGPT, Claude and Google AI Overviews, plus a technical AI-readiness scan of your site. It takes about 10 minutes; we'll email you the results.

Thorough analysis · ~10 min · No account required