Best AI visibility analytics for search optimization in 2026: what the numbers count, and what vendors document about them
Eight of the 18 AI visibility products we read publish how their headline percentage is calculated, and they use at least four different calculations.
We found that by reading the help centres, documentation and methodology pages of the 18 products named most often on the lists that rank for this question, on 12 October 2026, and putting the same five questions to each about its number. A 40% reading from one of those tools is not the same quantity as a 40% reading from another, and the documentation is where you find out which you are holding.
The same read turned up a second gap. Ten of the 18 say somewhere on their own pages that AI answers change from run to run. Only Profound publishes a measurement of how much its own number moves because of it.
That is how we think a "best" question about AI search reporting can be answered honestly: show what each number counts and what each vendor says about it. Nothing here is ranked, no winner is named and no price appears. Our interest needs saying early: we build a platform in this category. LiamVi is outside the 18 products we read, it is not counted, and no part of this page sets it beside anyone. The five questions are our choice of what matters in a number, shaped by what we built, so apply them to us too.
What makes AI visibility analytics "best" for search optimization?
No AI visibility analytics product is best in general.
The best one for your search optimization work is the platform whose number you can read, check and explain to someone else. Which platform that is depends on your buyers, your markets and which AI surfaces they use, so it is a decision we hand back to you.
The lists a marketer finds for this keyword do pick winners. We re-opened all eight pages in the results we captured on 1 September 2026 (the dated list is under Sources), and three of them carry a modified date after that capture. Several are written by a company whose own product sits first on its list; who writes these lists, and how much they agree, is the subject of our count of the recommended lists, so we do not repeat it here. Omnia's own page tells readers to treat "Rankings or scores (AI answers don't have ranks like traditional search)" as directional rather than exact (Omnia), and we agree with that line.
What the lists mostly compare, as SEO tool lists always have, is the feature set; what they leave out is the instrument behind each feature. A feature list tells you a tool tracks mentions on ChatGPT; it does not tell you how many times a day it asked, how it handled an answer that cited nobody, or what the percentage on the screen is a percentage of. That gap matters more in AI search than in traditional SEO, where a ranking report reads a results page that stays put long enough to check. Those are the facts that decide whether a trend line means something, and they live in documentation rather than on comparison pages.
What does an AI visibility number count?
An AI visibility number counts assistant answers, not web pages.
Each figure reflects whichever prompts a tool chose, put to whichever AI engines it covers, on whichever days it ran.
Three kinds of figure share the dashboard. A tracking figure (mentions, citations, share of voice) is a sample of answers. A content score or readiness rating is a model fitted to a target someone picked. A sentiment figure is a second model's judgement of how the first model described you. Each fails in its own way, and the rest of this page takes them in turn.
Three decisions sit under every tracking figure, and somebody made them for you: which prompts, how often, and on which engines. The prompts set the questions. The frequency sets how much of the engines' own churn you get to see. The engines set whose answers count. Change the prompt list and the number changes, with no change at all to your website. Replace ten prompts from a market that knows your brand with ten from a market that does not, and your share of voice drops against an unchanged competitor set in a week when your site stood still.
That is why the five questions below are about the number rather than the vendor: a careful vendor can still hand you a figure you cannot read if the documentation leaves out the part you need.
What did 18 vendors document about their own numbers?
Most of the 18 vendors we read document their prompts and their engines; far fewer document how often prompts run, how answers vary, or how the percentage is calculated.
On 12 October 2026, 16 named the AI surfaces they cover and 15 said where the tracked prompts come from, but only 10 stated how often the prompts run, 10 said anything about answers changing between runs, and 8 published how their headline percentage is calculated.
Here is the method, so you can repeat it or disagree with it. The eight ranking pages name 50 distinct products between them. We read the 18 named on at least three of the eight lists, because reading all 50 to the standard we wanted (every statement quoted from the vendor's own page and dated) was not possible in a single pass; the other 32 were not read and are not counted. SE Ranking and SE Visible come from the same company and count as two: the lists name them separately, and SE Visible "can be purchased as a standalone product" (SE Visible FAQ). For each product we opened its help centre, documentation or knowledge base first, then its product page where we found no documentation, then searched its own site for a methodology post on run-to-run variation. We fixed the five questions, and how each answer would be coded, before reading: where the prompt list comes from, how often prompts run, which surfaces are covered, what is said about variation between runs, and how the headline visibility number is computed.
Three limits travel with every count below. A "not found" means we did not find the answer on the pages we opened on 12 October 2026, not that the product lacks it; logged-in help centres were not opened. Documentation is not data quality: a platform that documents everything can still sample badly, and a quiet one can measure well. And the count does not rank anyone: there is no per-product total anywhere on this page, on purpose.
Where does the tracked prompt list come from?
Fifteen of the 18 products say where tracked prompts come from, and most draft them from your website or topics for you to edit.
Two build from keyword data instead. Semrush says it sources "billions of real prompts from AI search clickstream data and Google's keyword dataset for AI Overviews" (Semrush knowledge base). Ahrefs says its Brand Radar prompts are "modeled after real keywords from our database, asked by real humans", expanded through People Also Ask and what it calls semantic fanout (Ahrefs help centre; Ahrefs methodology).
The generated kind is more common. Scrunch tells new users to review "the 5 starter prompts that Scrunch automatically generated for you, then add 5-15 more" (Scrunch help centre). Conductor "will generate up to 100 prompts for a topic" (Conductor documentation). AthenaHQ, BrightEdge, Nightwatch, Peec AI, Profound, SE Visible and Writesonic describe prompts drafted for you on their own pages, in most cases with room to add your own. AirOps (AirOps University), Rankscale (Rankscale help centre) and SE Ranking (tracker page) track the prompts or terms you enter yourself.
Otterly.AI states the underlying limit plainly: "AI search engines like ChatGPT and Perplexity do not publish query data. There is no public keyword tool for AI search; at least not yet" (Otterly.AI help centre). Every prompt list is therefore an estimate of what people ask, built by somebody, and the number inherits that estimate. That holds for AEO and GEO dashboards alike: the label on the product does not change where the questions came from. For the three products where we found no answer (Brandlight, Hall and Rankability), the keyword list behind the figure is the first thing to ask about. Otterly.AI calls the whole feature prompt monitoring, and the name fits: what is monitored is the list.
How often do the prompts run?
Ten of the 18 state how often prompts run, and the cadences they state range from hourly to monthly.
Two tools can therefore look at the same week with very different amounts of data behind each point.
Five say daily, at least by default. Profound: "Prompts refresh once per day on Eastern Time" (Profound help centre). Otterly.AI: it "monitors your brand visibility across all major AI search engines on a daily basis", and "all prompts run on the same daily cadence" (Otterly.AI help centre). Peec AI runs prompts daily, in a sentence we quote in full under variation below. Semrush's prompt tracking is "updated daily", while its Brand Performance reports are "updated weekly" (Semrush knowledge base). SE Visible checks "Daily by default", with a weekly option (SE Ranking knowledge base).
The other five describe something less uniform. Scrunch collects "daily for the first 14 days", then "shifts to a default 72-hour refresh cadence" (Scrunch FAQ). Ahrefs runs two instruments. Its main prompt dataset refreshes chatbot data "monthly, using a 90-day reporting window for stability and consistency", while Google AI Overviews and AI Mode "update continuously" (Ahrefs methodology); its custom tracked prompts run on a schedule you set: "Tracked prompts can be updated monthly, weekly, or daily" (Ahrefs help centre). Conductor lets you choose daily, weekly or monthly per topic and recommends weekly as the starting point (Conductor learning library). Rankscale's scheduling feature offers hourly, daily, weekly and monthly per tracked term (Rankscale help centre). AthenaHQ lets you set "a weekly cadence, including the selected days and a repeat interval of one to four weeks" (AthenaHQ documentation).
Monitoring cadence is part of a number's label, as a sample size is part of a poll's. A daily reading is built to catch a change quickly, after a content update say; a monthly reading over a 90-day window is built to sit still. Each suits its own job. Mixing them in one trend line, or comparing two tools on different schedules, treats two instruments as interchangeable. Rankscale's help page warns its own customers about the same thing: mixing hourly and weekly intervals in a single topic "can produce uneven data points that make trends harder to interpret" (Rankscale help centre).
Which AI surfaces does each number cover?
Sixteen of the 18 name the AI surfaces they track, and the lists run from four engines to eleven.
A single AI visibility figure can therefore stand for very different slices of where your buyers ask.
At the narrow end, Nightwatch's documentation lists "ChatGPT, Perplexity, Google AI Overview, Google AI Mode" as providers (Nightwatch documentation). At the wide end, AthenaHQ's setup screen offers "ChatGPT, AI Overview, Perplexity, Gemini, Copilot, AI Mode, Grok, Claude, DeepSeek, Meta AI, Mistral" (AthenaHQ documentation), and Rankability lists eleven as well, with the note that "Available platforms depend on your plan and add-ons" (Rankability). Scrunch names nine (platforms FAQ). Ahrefs adds a caveat worth knowing: Grok data is paused because "we are temporarily unable to gather new data", and Claude is tracked "via API" for custom prompts only (Ahrefs help centre).
Nearly every list includes ChatGPT and Google AI Overviews, and Profound's documentation calls each such surface an answer engine (Profound help centre). Where the lists part company is Claude, Copilot and Meta AI. If your buyers ask Gemini or Perplexity more often than ChatGPT, a figure built mostly on ChatGPT answers describes somebody else's market; that judgement is yours, because only you know where your buyers ask. Our own limit, stated at the claim: we measure mostly ChatGPT, sample Gemini and Perplexity lightly, and have barely looked at Claude or at a Google AI Overview.
What do vendors say about answers changing between runs?
Ten of the 18 say the same prompt can return a different answer on a different run, and only Profound measures what that does to its own number.
Profound's post on whether a daily run is enough answers the question with a measurement. It ran two copies of a single setup side by side for two weeks, "the same 753 prompts across the same 7 platforms" in the US, the first once a day and the second ten times a day. Its finding, in its words: "Once a day already lands within about 2 percentage points of a 10×-a-day reading for visibility." It also names the part no run count fixes: "You can't measure your way past drift" (Profound blog). That is a vendor measuring its own instrument and publishing the result with its sample and method. The limit is the vendor's own: a single portfolio, the US, two weeks.
The other nine state the fact without a figure. Otterly.AI: "AI search responses aren't deterministic; the same prompt can return different results on different runs" (Otterly.AI help centre). Peec AI runs prompts "daily, then analyzes patterns over time, since AI responses naturally vary day to day" (Peec AI documentation). Semrush: responses are "fast-changing and highly personalized, which means no platform can provide exact numbers on visibility" (Semrush knowledge base). AirOps: "your brand can appear in one run and vanish in the next" (AirOps glossary). SE Ranking describes a handling method: it "parses each tracked prompt three times" (SE Ranking blog). Ahrefs, Conductor, Scrunch (blog post) and SE Visible (blog post) say much the same on their own pages, and Conductor adds a feature that "runs a variance analysis across your tracked topics" to show where the data has settled (Conductor academy).
Our harvest showed the same behaviour: an identical question, put twice to the same assistant, could come back with a different answer citing different pages. We have not put a number on that for our own sweeps, so here Profound has published a measurement we have not.
What is the visibility percentage a percentage of?
Eight of the 18 publish how their headline number is calculated, using at least four different constructions.
Two tools showing 40% for the same brand may therefore be reporting different quantities.
The plainest construction is a share of all answers. Peec AI: "Visibility Score = (Number of responses mentioning your brand / Total responses) × 100" (Peec AI documentation). Otterly.AI's per-prompt figure is "The share of AI responses for this prompt that mention your brand" (Otterly.AI help centre), and Writesonic's is the "Percentage of tracked AI answers where your brand is mentioned" (Writesonic documentation).
Profound's denominator is narrower: "the number of responses that include your brand divided by the total number of responses that include at least one brand" (Profound help centre). An AI answer that names no brand at all drops out of the bottom of that fraction. On the same set of answers, that construction gives a higher figure than a share of all answers whenever some answers name nobody.
Two vendors weight by position. SE Visible (SE Visible FAQ) gives each mention a weight of "(total_brands − best_position + 1) / total_brands" and then averages over "every response in scope, including those where the brand was not mentioned." Rankscale multiplies a detection rate by a position discount, "(appearances / total runs) x (1 / (1 + 0.1 x (average position - 1))) x 100", and gives a worked case: a brand found "in 80 of 100 runs at an average position of 2 scores about 72.7%" (Rankscale help centre).
Ahrefs weights its Estimated Impressions by how often a query is searched on Google, and says exactly what that is: "This is a modeling choice, not a measured relationship: we don't claim a validated link between Google search volume and how often a query is asked inside an AI tool" (Ahrefs methodology). Semrush describes its AI Visibility rating against the median mentions of a competitor set it identifies automatically, and its data page names two components without publishing how they combine (Semrush knowledge base; Semrush data page). Conductor explains why a share of voice runs lower than a mention rate: share of voice divides by every competitor and brand mention, "making the total denominator much higher than the number of prompts you are reporting on" (Conductor documentation).
The practical consequence is simple. Before you compare two tools, or a tool against its own history after a change of plan, find the formula. Where the documentation gives none, the number is a reading whose units you have not been told.
What insight does a tracking figure actually support?
A tracking figure gives sound insight on direction, empty prompts and repeat competitors, and poor insight on cause.
Direction is the safest of these: on a prompt list and a cadence that both stayed fixed, several weeks of movement support trend analysis, while a single week rarely does. Empty prompts come next, because a prompt that keeps returning nothing tells you about the question, not about your pages. Repeat competitors are the third, and there the insight is not the tally but what that competitor's cited pages contain, which you learn by opening them. Cause is where a tracking figure fails. A sampled count cannot link last month's article to this month's mentions, because the engines changed, the prompt list may have changed, and run-to-run variation sits beneath both.
What the count does not tell you
The count records what is written down on the pages we could open, and nothing about how well any of these products measures.
Two products show the limit. Hall appears on three of the eight lists. Its own site, opened on 12 October 2026, reads "Hall is now part of Tracksuit" (Hall), dated 21 July 2026, with a waitlist in place of a product page, so there was no documentation to read and Hall is counted as not found on all five questions. AthenaHQ's documentation describes a glossary that explains "how specific formulas are calculated" (AthenaHQ glossary), but the same page says it renders blank for anyone without an active subscription, so its formula is probably documented for customers and was not readable by us.
Several of these vendors probably document more about their AI search monitoring inside logged-in help centres than in public. What the public read measures is how much a buyer can check about a platform before that call, which is a reasonable thing to know and a narrow thing too. Every quote is dated because a vendor page can change the day after we read it.
Why does the framing of a tracked prompt decide whether a citation exists?
A prompt's framing can decide whether the engine searches the web at all.
When nothing is looked up, no source is cited, and a citation count shows a zero that says nothing about your content.
We checked this on ourselves first, writing the prediction down before the run and holding the topic constant while only the wording varied. The outcome is recorded on our research page: "none of the conceptual “what is X” framings triggered source retrieval on the engine we measured, while every dated, comparison-shaped framing did." The page states its limits beside the result: five topics, mostly one engine, and engines differ. The direction held; the rates do not transfer.
The platforms say the same in their own documents. Google's page on grounding with Google Search, last updated 10 October 2026, says that "there are several reasons this metadata might not be provided, and the prompt response won't be grounded" (Google Cloud documentation). OpenAI's announcement of ChatGPT search approaches it from the user's side: "ChatGPT will choose to search the web based on what you ask" (OpenAI). Conductor documents the same effect from the tracking side: a tracked prompt can be submitted with "no AI response" returned, which is "most common for Google AI Overviews, where AI responses are generated only for 50–70% of prompts" (Conductor documentation). An answer given with nothing looked up can still name brands: OpenAI's help centre says that with web access off, "Responses are based on what the model has learned during training" (OpenAI Help Center), so a mention rate can move on a prompt where no page could have been cited.
So an empty citation cell has two possible causes. Either the engine searched and cited someone else, or it never searched. Only the first is a problem new content can fix. Read your tracked prompt list as part of the instrument: how many prompts are conceptual "what is" questions, and how many name a product, a comparison, a year or a market? A list heavy with the first kind produces citation zeros that are facts about the prompts, not about your content.
Why can two careful tools show different numbers for the same brand?
Two careful tools disagree because they ask different prompts, on different schedules and engines, and divide by different denominators.
All of that comes before any question of accuracy.
Sampling a system that moves, at different moments, two tools will land apart even when each is accurate, and with at least four constructions in use the gap between two dashboards proves nothing about either. A side-by-side trial of two monitoring products measures how far apart they sample. That is worth knowing; it is not the same as knowing which is right.
Whether a given vendor's data holds up is a separate test with its own method, set out on our page on checking a platform's data accuracy, where our own cross-check of Surfer's citation data is reported; Surfer's numbers held up.
What is a content score measured against?
A content score is measured against whatever target its maker picked; on our citability checklist, the 43 pages that ChatGPT cited across our harvest have a median of 65.
Those 43 are articles of 800 words or more, and only around one in twenty of them gets to 85, so we read a content score as a floor, never a finish line.
Both figures, with their samples, are on our research page, alongside a third: professionally published articles have a median near 76 on our SEO-terms meter (44 articles measured). That is why our bands label 55 to 84 as strong and treat 85 or more as information rather than a goal. The label describes where those cited and professional pages sit; we have not validated it as a threshold for being cited. Whether pushing past the cited pages' range produces padded content is our reading, not a result: we have not tested whether a higher reading costs a citation, and we cannot show that adding what cited pages share causes one.
On a dashboard a content score and a tracking figure often look alike, but they are different objects. The tracking figure samples what engines did; the meter is a model fitted to a target that is seldom printed next to it. Asking what it was fitted to is fair, and the question has an answer. A coverage checklist shows what is missing, which is useful; it does not grade writing, and no score, ours or any vendor's, is the probability that a page gets cited.
We know the difference because we got it wrong. Our first meter rewarded copying the vocabulary of pages already ranking; compared with independent cold reads of the same articles, it disagreed with them, and we rebuilt it as a checklist of what is missing. Those cold reads were done by automated reviewers rather than people, and our research page says so in its own words: "independent reviewer agents on a seven-dimension scorecard, not yet by a person". It also reports that a person's blind read of those articles is "prepared and unfinished", and we will publish its result either way.
What about sentiment and blended performance scores?
Sentiment is a model's judgement of another model's tone, and it inherits every limit of the tracking figure.
The prompts, the schedule and the engines all carry over, with a judgement call on top.
None of that makes sentiment useless: if forty answers call your product expensive, you want to know, and the cause may be content you can fix. But we have not measured sentiment accuracy anywhere, in our product or any other, so we have no figure, and we treat a sentiment trend as a reason to go and read the answers. A blended performance figure that folds mentions, citations, position and sentiment together into one headline number adds the vendor's weights on top. SE Visible's FAQ shows the other approach: its visibility score leaves citations out, and sentiment "has its own formula and does not affect the Visibility score" (SE Visible FAQ). Ask for the weights, and for what the total does if an input is missing.
What is outside every AI visibility number?
A page that no engine went looking for is outside every count on the dashboard.
A citation figure cannot credit that page, and a mention drawn from what a model already learned points to no page at all. Nor can a traditional SEO rank-tracking report stand in for either: in our harvest, ranked pages and cited pages barely overlapped.
Our research page gives the counts for that 33-keyword harvest: 276 pages ranked, ChatGPT cited 73 pages for the same queries, and only 9 pages were on both lists: 9 of 276 on the ranking side, 9 of 73 on the cited side. In the page's words, "Seven citations in eight came from outside the first page of results." Its stated limits apply here too: three related industries, mostly ChatGPT, counts of pages rather than domains, a single point in time. Read the counts as what we observed, not a law. A strong report from your SEO tools says nothing about whether assistants quote you, and AI visibility tracking that merges the two into a single headline hides which half moved.
Share of voice adds a blind spot of its own: it is a ratio against a competitor set whose denominator you rarely see. It shifts when a competitor's mentions shift, when the prompt list is edited, and when answers that cited nobody are treated another way. Some tools drop those answers as no data; others count them as a competitor's gain. Both choices can be defended, and from identical observations they draw different trend lines. Ask your tool which it makes, and how many of your tracked prompts fell into that bucket last month.
A first-party console is a different instrument again. Microsoft describes Bing Webmaster Tools' AI Performance report as "a new set of insights that shows how publisher content appears across Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations" (Bing Webmaster blog): Microsoft's own surfaces and nobody else's. Our reading of that report sets out what it counts.
Which five questions should you put to a vendor?
Five questions separate a number you can read from a number you cannot, whatever feature list sits around it, and they are the same five we put to the 18 products above.
- Who wrote the prompt list behind this figure, and can I see it? The list itself, not a count of it.
- How often do the prompts run, and on which days? A trend whose cadence changed halfway through is really two trends.
- Which AI surfaces does the figure cover, and is any of them paused or tied to a plan?
- How far does the figure move between runs when your site has not changed? A vendor that has measured it can say; a vendor that has not should say that instead.
- What is the percentage a percentage of? Ask for the formula, and for what it does with an answer that names no brand.
Our own answers, as an AI visibility tool maker, sit outside the count and are offered so you can hold us to the same standard. Our prompt list is the set of keywords the customer chooses to track, so the person reading the number built part of the instrument. Our sweeps run weekly. We have observed run-to-run variance but have not published a figure for it, which is a gap in our answer to question four. Our checklist's target is described above, with the median of 65 across the 43 cited pages beside it.
Passing all five does not make a figure right. It makes it checkable, and a checkable figure is a measurement where an uncheckable one is only a claim.
How can you check your own dashboard this week?
Five tracked prompts, two engines, two runs and two days give you 40 observations of your own.
Four observations per prompt and engine is a spot check, not a variance study, and it is still more than most dashboards disclose about their own movement.
Budget an hour or so over the week; no paid tools are needed. Choose the two engines your buyers actually use, whatever your platform tracks, and if a Google AI Overview matters more to your market than a chatbot answer, use an ordinary Google query as the second surface. Copy each prompt exactly as your AI visibility tool stores it, and note three things per run: whether any source was cited, whether you appeared, and which competitor was named instead.
Say the limit now rather than later: an app you are logged into is a different instrument from a tool's scheduled run. Ahrefs says it asks its prompts "without any stored user data or prior context" (Ahrefs help centre), and Peec AI says its collection reflects "the reality for a logged-out user" (Peec AI documentation). Your account brings its own history, so your notes measure your instrument and do not replace theirs.
Then set your notes beside the tool's report for the same days. A prompt with no sources on either day had no citation slot when you checked; find out how the tool scored it. A prompt whose sources changed between days shows variation you watched happen: too few observations for a variance figure, enough to stop you treating a small weekly move as a result. If you and the tool disagree on which way the week went, the direction rather than the figure, raise it with support before you think about switching.
If the two agree in broad terms and some movement shows, you have seen that your number shifts for reasons unconnected to anything you published, though not by how much: that takes far more runs, on the scale of Profound's test above. The pattern worth querying is a report that never moves while your own runs keep disagreeing.
What people ask us
Which AI tool offers the best visibility optimization? We do not name a winner, and we would not trust a page that did without showing its method. The tool that fits covers the AI surfaces your buyers use, at a cadence that suits how fast your market moves, with a formula you can read. On the pages we opened, 8 of the 18 most-listed products publish that formula; the best first step is to start there and ask the others for theirs.
What is the best AI for search engine optimization? It depends which question you mean. Which assistant to write with is something we have not measured. Which engine to optimise for has a plainer answer: the ones your buyers use, because ChatGPT, Gemini, Perplexity and Claude behave differently. Our data is mostly ChatGPT, and the clearest thing in it is that a question's framing settles whether an engine consults the web at all.
Where can you find the best LLM optimization for AI visibility? In AI search, generative engine optimization (GEO), answer engine optimization (AEO) and LLM optimization name the same work you can actually observe: which questions set off retrieval, which pages are quoted, which sentences are lifted. Nobody outside the labs can see inside a model, so improving what a model can retrieve is genuine work, while improving the model itself is not something any vendor can show you evidence for.
Is SEO dead now with AI? No. Our harvest is the reason we can say it without drama: ranked pages and cited pages were close to separate sets, 9 pages shared between 276 ranked and 73 cited. Two channels exist side by side. Traditional SEO keeps earning the traffic it earns, and the content work behind it still works; AI search is a second channel, measured another way.
Is AI visibility analytics a single product, or a stack? For most teams, a stack. None of the products we read covers every AI search surface at every cadence, and rank tracking watches a different set of pages altogether. Pick for the surfaces your buyers use on whichever AI platform they prefer, keep rank tracking where rankings still bring traffic, and read any blended figure as a summary.
The first four questions above are the ones Google showed searchers asking about this keyword in the results we captured (United States, 1 September 2026), kept close to their wording; the fifth is ours.
What we cannot tell you
The documentation count cannot tell you which product measures best.
It covers 18 of the 50 products on eight ranking pages, read from public pages on a single day, and it cannot tell you what any vendor documents behind a login or what changed on a vendor's site after 12 October 2026.
Our citation data is small and early: 33 keywords, three related industries, mostly ChatGPT, and every claim drawn from it carries that limit. In full, 241 articles were scored, 43 cited and 198 ignored, plus 21 matched pairs; only articles of 800 words or more were scored, so that set is smaller than the harvest's 73 cited pages. That is enough to shape our own product decisions and too little to state laws. We have not shown causation or measured sentiment accuracy, and our own variance figure is still unpublished.
The dataset grows with every weekly sweep. If a conclusion on this page stops holding, the documentation count included, we will correct it publicly and say what changed.
Sources
- LiamVi — Research: our methods, our samples — the 43 cited pages and the median of 65, the SEO-terms meter across 44, the framing ladder, the 276 / 73 / 9 overlap, the dataset arithmetic and the published limits. Page last updated 12 August 2026; checked 12 October 2026.
- LiamVi — checking a platform's data accuracy — run-to-run variance in our harvest and the Surfer cross-check. Checked 12 October 2026.
- LiamVi — how much the recommended lists agree, and who wrote them — the count of the recommended lists. Checked 12 October 2026.
- LiamVi — Bing Webmaster Tools AI performance report — what a first-party console counts. Checked 12 October 2026.
- Microsoft Bing — Introducing AI Performance in Bing Webmaster Tools — what the report covers, in Microsoft's words. February 2026; checked 12 October 2026.
- Google Cloud — grounding documentation — grounding metadata may not be provided. Last updated 10 October 2026; checked 12 October 2026.
- OpenAI Help Center — answers without web access — with web access off, responses rest on what the model learned in training. Updated about two months before our check; checked 13 October 2026.
- OpenAI — announcement of 31 October 2024 — the assistant decides when to use the web, based on what you ask. Updated 5 February 2025; checked 12 October 2026.
The ranking pages we read — the results captured for this page's keyword on 1 September 2026 (United States), each re-opened on 12 October 2026; dates are each page's own: - Omnia — its list of 26 tools — modified 7 October 2026. - 42DM — its top 10 and 21 others — modified 4 December 2025. - Nick Lafferty — nine platforms ranked by AEO score — modified 20 August 2026. - Profound — 18 tools for marketing agencies — undated. - WP Engine — seven tracking tools — modified 17 September 2026. - NoGood — nine optimization tools — modified 14 August 2026. - Cognizo — nine platforms — undated. - Analytify — six tools — modified 30 September 2026.
The vendor pages we read, all checked 12 October 2026: - Ahrefs — What is Brand Radar and how to use it, Brand Radar methodology and custom prompts (the custom prompts page checked 13 October 2026). - AirOps — glossary entry and AirOps University tutorials. - AthenaHQ — Setting up your prompts, Add website and Glossary. - BrightEdge — AI Catalyst. - Brandlight — home page. - Conductor — FAQs, tracking best practices and prompt generation for AEO. - Hall — home page. - Nightwatch — AI LLM tracking. - Otterly.AI — monitoring interval, relevant prompts and prompt monitoring. - Peec AI — Welcome to Peec AI, Visibility and Setting up your prompts. - Profound — Answer Engine Insights overview, Create, manage and tag prompts and Is once a day enough?. - Rankability — AI tracker. - Rankscale — Automatic updates and Metrics in Rankscale. - Scrunch — data collection FAQ, platforms FAQ, quick-start guide and monitoring questions answered. - SE Ranking — tracker page and a blog post on run-to-run change. - SE Visible — FAQ and How SE Visible works. - Semrush — toolkit overview, AI SEO metrics and Where the data comes from. - Writesonic — GEO: getting started and Dashboard.
