AI Visibility Tools Compared by Where Their Data Comes From

AI visibility tools disagree about the same brand because they run different query sets. Four of the ten external platforms name a real-conversation dataset. The other six build their sets from keywords, models, or questions supplied by the client.

No major AI platform shares its query data, so every vendor estimates differently, as Ahrefs explains in its AI-adjusted volume methodology. Vendor methodology was checked on 31 August 2026. Where a source or method is not published, the relevant entry says so.

Disclosure: Klarivo competes in this category. Klarivo Monitor is scored from published material under the same rubric and finishes last. The scores measure what each vendor discloses about its method, not how well its product works.

Key Takeaways

• Query-Set Provenance: the biggest driver of any AI visibility number is which prompts get run. Two tools with identical engine coverage will disagree if their prompt sets differ.
• Capture Path: reading the consumer interface and calling a provider API return different answers. Writesonic publishes a side-by-side showing the API pulling far fewer sources.
• Not Published Is a Finding: a blank cell below is a real result. It names the question to put in writing before you sign.
• Contrarian: four of the ten name a real-conversation dataset. Only two say whose conversations they are. The other two name a category and stop.
• The Gap Is Your Category's: how far two tools diverge depends on how differently your buyers talk from a keyword index. In gaming that gap is wide. In real estate technology it is narrow.

What an AI Visibility Tool Reports, and What It Infers

Every AI visibility tool presents three classes of data, and dashboard labels rarely separate them. Collected data includes raw answers, brand mentions, and cited URLs captured during each run. Calculated metrics include mention rate, position, share of voice, and sentiment scores derived from those answers. Modeled estimates include prompt volume, topic volume, and anything presented as demand.

All three can appear in the same table. That is how a buyer ends up treating an estimate as a measurement.

Why This Article Is Not the Other Two

Klarivo has published two comparisons of this field already. One scores seventeen tools against a shared rubric. The other ranks them by coverage tested rather than claimed.

Neither asks where the numbers start. This one asks only that, so features appear below only where they bear on provenance.

The mechanics of scoring position and sentiment on a single engine live on their own page.

Where Each Tool's Prompt Data Comes From

The table is the argument. Two numeric columns carry the score and the published scale, and the two provenance columns carry the verdict.

Cells read "Not published" where the vendor's own pages do not answer the question.

Eight of the eleven rows carry at least one “Not published” entry, and seven of those appear in the dataset-scale column. Four vendors publish a numeric scale: Semrush, Profound, Ahrefs, and Writesonic. A blank is the question you take into the contract.

How These Were Scored

The rubric measures disclosure, not accuracy. A tool can measure well and publish little, and this page cannot tell the two apart. It can tell you which AI visibility tools let you argue with their numbers.

Published tiebreakers, in order. A named source beats an unnamed category. A published refusal beats an unpublished estimate.

A dated page breaks any remaining tie.

The disclosed bias: we weigh query-set provenance heaviest, because everything downstream inherits it.

Anyone running AI search tracking can repeat the underlying research in an afternoon. Open each vendor's docs, find the page answering "where do the prompts come from", and write down what it says.

Every vendor answers that one. Three of the eleven never say how the answer itself is captured.

Synthetic Prompt Sets Against Sampled Real Queries

Ahrefs names four routes the category takes to a volume figure. Extrapolate from opt-in browser panels, generate variations of seed keywords and assign volumes with no measurement, derive from traditional search volume, or decline to estimate.

Rankscale supplies a fifth, reconstructing prompt clusters from the models themselves.

Those five describe the query set. The capture path is a separate question, and most buyers never ask it.

None of the five is inherently wrong. A user-written set measures your hypothesis about the market. A sampled set measures the behavior represented in its sample and ignores your hypothesis.

Buyers comparing AI visibility monitoring tools on a feature grid miss that these are not one product.

The Ten External AI Visibility Tools, Compared by Data Source

The published total sets the published order. Every entry carries the same five fields, including the publisher's own platform in the final section. Shortlisting best AI visibility tools or best AI search monitoring tools from a feature grid hides all of it, so read the Verdict lines first.

Semrush AI Visibility Toolkit (Best for the Most Complete Published Method)

Its knowledge base page on AI visibility data is the most complete methodology document in this field. It also explains why it refuses to publish a per-prompt volume.

Query set: 317 million prompts and responses across ChatGPT, Gemini, AI Overviews, and AI Mode, sourced from AI search clickstream plus Google's keyword dataset. Updated daily across 117 regional databases.

Volume method: topic level only. It combines third-party data on real AI interactions with its own models, and states that individual prompts are too specific to measure directly.

Capture path: published. "Prompt responses are captured from real requests and not via any APIs of LLMs." Prompt Tracking runs your own prompts daily.

Stated limits: the page says no platform can provide exact numbers, and calls its own metrics directional signals.

Score: 25 query set, 20 volume, 20 capture, 15 limits, 12 reproducibility. 92/100.

Verdict: shortlist it for the most auditable method here, especially if you already pay for the platform. Look elsewhere if you need the clickstream provider named, because it is not.

Profound (Best for a Named Panel Behind the Prompt Set)

Its Prompt Volumes page names the data source rather than gesturing at it. Almost nobody else in this table does.

Query set: three routes, all published. Auto-generate from your brand configuration, upload your own, or pull from Prompt Volumes, described as real prompts submitted by actual users.

Volume method: licensed conversations from multiple double opt-in consumer panels, anonymized and scrubbed of personal data, then modeled to correct demographic and geographic bias. Coverage runs from January 2025, refreshed weekly.

Capture path: published. Responses come from the browser front end rather than the API, and every tracked prompt runs daily.

Stated limits: it states that figures are modeled as well as observed, and names ten covered countries. It publishes no "not useful for" list.

Score: 30 query set, 20 volume, 20 capture, 10 limits, 10 reproducibility. 90/100.

Verdict: shortlist it if the panel is the thing you are buying and the budget assumes a sales cycle. Think twice if you sell outside those ten countries. Coverage is where it thins out.

Ahrefs Brand Radar (Best for Arithmetic You Can Re-Derive)

Its own FAQ states that Brand Radar runs large sets of prompts and records whether your brand appears. The volume attached to those prompts is derived, and the derivation is published in full.

Query set: two sources, both named. A pre-built library of more than 400 million search-backed prompts, converted from real queries in Ahrefs' own search index into conversational questions, plus Custom Prompts you define. The library is search demand rather than an AI conversation log, and the page says so.

Volume method: AI-adjusted volume. Each prompt's parent keyword is looked up in Google search volume, then multiplied by a platform ratio drawn from Ahrefs' AI referral traffic against Google organic. A reader with a seat can reproduce it.

Capture path: not published. The FAQ calls it "a sampling-based approach, not real-time monitoring", without saying whether the answer is read from the interface or an API.

Stated limits: the strongest here. The post names four uses the metric suits and three it does not, including AEO key performance indicators. The FAQ adds that results are "a structured sample, not every mention your brand receives".

Score: 24 query set, 20 volume, 12 capture, 15 limits, 15 reproducibility. 86/100.

Verdict: shortlist it if you already run Ahrefs and want a benchmark you can explain to a skeptic. Skip it if your buyers phrase things nothing like they search, because the derivation assumes they roughly do.

Scrunch (Best for Publishing What It Will Not Estimate)

Its FAQ on how it estimates AI search volume does something no other vendor here does. It publishes a refusal.

Query set: you define the tracked prompts. Topic trends come from a large-scale panel dataset supplied by third-party providers.

Volume method: topic level only, from panel data plus proprietary modeling. The page states that prompt-level data is not yet reliable enough to measure, so it does not offer it.

Capture path: published. Browser automation and official platform APIs, chosen per platform, measured against a continually updated response dataset.

Stated limits: stated in its own words. "Data is directional, not exact."

Score: 20 query set, 18 volume, 20 capture, 15 limits, 12 reproducibility. 85/100.

Verdict: shortlist it if "we cannot measure that yet" reads as maturity. Skip it if your board wants a prompt-level demand figure. It is now a Sitecore company and serves from scrunch.com.

Otterly.AI (Best for Labelling an Estimate as an Estimate)

Its prompt research page carries the most honest sentence in this comparison. OpenAI does not disclose prompt volume, so Otterly built an Estimated Intent Score to give direction instead.

Query set: you define the prompts. Its research tool generates candidates from a seed topic, competitor data, live engine behavior, and Google Search Console.

Volume method: none claimed. The Estimated Intent Score is named as an estimate on the page that sells it.

Capture path: published. It queries seven engines daily as a neutral, non-personalized user across 65 countries or more, so browsing history does not skew results.

Stated limits: it names the disclosure gap directly rather than working around it.

Score: 22 query set, 12 volume, 20 capture, 12 limits, 10 reproducibility. 76/100.

Verdict: shortlist it for a defined prompt set run cleanly, with no invented demand figure. Think twice if you need discovery, because the set still starts with you.

Rankscale (Best for a Named, Attributable Reconstruction Method)

Its entity facts page publishes a method with an author's name attached. That is unusual enough to credit.

Query set: Prompt Decoding, an exclusively licensed method developed by Hanns Kronenberg. It reconstructs representative prompt clusters from large language models, imputing user intent where direct logs do not exist. You can also define terms yourself.

Volume method: no separate volume figure is published beyond that reconstruction.

Capture path: published, and unusually specific. It separates seven AI search interfaces from thirteen named model engines, with tracking intervals from bi-hourly to bi-monthly.

Stated limits: it claims congruence with an external 2025 working paper rather than publishing its own error bars.

Score: 22 query set, 10 volume, 18 capture, 8 limits, 12 reproducibility. 70/100.

Verdict: shortlist it if you want interface and model results separated rather than blended. Skip it if a reconstruction from the model is the one thing you were trying to avoid buying.

Peec AI (Best for Per-Chat Location and Source Detail)

Its documentation on how each tracked chat is produced is the clearest per-response record here. It includes caveats that would be easier to leave out.

Query set: you define the prompts, with a prompt builder and discovery suggestions available.

Volume method: a Prompt Volume score of 1 to 5, based on real-time search trends, AI conversation data, and industry signals. Those are categories rather than named sources, and the output is a band, not a count.

Capture path: published. Chats are the responses produced by running your prompts daily, each stamped with the location prompted from. Gemini can only be prompted from the United States for now.

Stated limits: it separates sources from citations, and warns that model source selection is partly random.

Score: 20 query set, 8 volume, 18 capture, 12 limits, 10 reproducibility. 68/100.

Verdict: shortlist it if per-chat forensics matter more than a demand number. Skip it if you need Gemini results outside the United States, because the docs say you cannot.

Writesonic (Best for a Capture Path Documented in Full)

Its documentation on why it reads interfaces rather than APIs publishes a side-by-side comparison of the two, making it the most useful capture-path document here.

Query set: generated from multiple named inputs. Writesonic says Prompt Explorer draws on more than 120 million AI conversations across platforms, public discussions on Reddit, Quora, and Stack Overflow, enterprise partner data, and industry-specific usage trends.

Volume method: published at a high level. Writesonic describes AI Search Volume as an estimate produced by a multi-source ensemble system. It names the categories of data behind the estimate but does not publish the complete formula or identify every underlying provider.

Capture path: published in detail. It reads chatgpt.com, Google AI Overviews, and perplexity.ai directly, simulates locations, and shows the API returning three to five sources where the interface returned fifteen to twenty.

Stated limits: the documentation identifies AI Search Volume as an estimate and discusses platform volatility, but it does not publish error margins or fully define the limits of its underlying conversation sample.

Score: 25 query set, 12 volume, 20 capture, 8 limits, 8 reproducibility. 73/100.

Verdict: shortlist it if capture fidelity across markets and a broad discovery dataset matter. The data pool is wider than one product’s users, but the underlying conversation and partner providers are not fully named.

SE Visible (Best for a Published Scoring Formula)

Its product FAQ publishes the arithmetic behind the visibility number, worked example included. Nobody else in this table does that, including us.

Query set: you define it. No discovery dataset is published.

Volume method: none published at all. That costs points here and misleads nobody.

Capture path: published. It collects real AI responses across five systems, stating they are not simulations or API shortcuts, and aggregates three runs per prompt weekly.

Stated limits: citations are excluded from the visibility score and tracked separately, so the number covers mentions and position only.

Score: 18 query set, 0 volume, 20 capture, 12 limits, 15 reproducibility. 65/100.

Verdict: shortlist it if you need to defend a number line by line to a finance team. Skip it if you wanted demand data, because it does not pretend to have any.

AthenaHQ (Best for Prompt Discovery From Your Own Search Console)

Its prompt setup documentation is the most practical discovery guide here. It also carries the widest gap between what a method claims and what it names.

Query set: four routes, all documented. Generate from your brand, type them in, import a CSV, or run Discover across your website, Google Search Console keywords, Reddit and YouTube discussions, and competitor keyword gaps.

Volume method: a Query Volume Estimation Model, drawing on 50 or more sources across public, partner, and proprietary data. None of those sources is named.

Capture path: not published on the pages read.

Stated limits: it publishes a 95 percent accuracy claim without naming the validation dataset. That is the opposite of a stated limit.

Score: 20 query set, 10 volume, 5 capture, 5 limits, 5 reproducibility. 45/100.

Verdict: shortlist it if the social and Search Console discovery routes solve a real problem, because they are genuinely good. Ask for the QVEM sources in writing first, because the accuracy figure has nothing published behind it.

What the Numbers Mean When the Data Source Changes

The size of the disagreement is a property of your category, not of the tools. Two vendors land close together where buyer language resembles a keyword index, and far apart where it does not.

Two hypothetical examples show how the gap can widen or narrow.

Gaming illustrates the wide case. A player asks an engine what to play after a role-playing game they loved but whose combat they disliked. A prompt set derived from search volume returns "best RPG games 2026" instead, because that phrase has a parent keyword and a monthly figure behind it.

The two sets barely intersect. Two AI visibility tools then report different numbers for the same studio, and both are right about their own set.

Real estate technology illustrates the narrow case. Procurement teams ask in category language, close to what they would type into Google, because the software category names are the words they already use internally.

A derived prompt set tracks that closely, and the gap between the two methods shrinks to something a buyer can ignore.

The reading rule that falls out of this is the article's own, and you can test it in an hour. Write down ten questions your last ten buyers actually asked.

If most survive being turned into a two-word search, a derived method will serve you. If most do not, buy a sampled one or accept that you are measuring your own guess.

Choosing by Provenance Rather Than Feature Count

Provenance sounds like a technicality until two numbers disagree in a board meeting. Any AI visibility platform can produce a chart. Far fewer can tell you which questions produced it.

That is the difference between a number you can defend and one you can only repeat.

"I need to prove this channel matters." Buy a sampled method and put the source in the footnote. Semrush and Profound both name a dataset. Do not present a derived figure as market size, which Ahrefs itself warns against.

"I need to find prompts I have not thought of." Buy discovery, because a user-written set by definition cannot supply it. AthenaHQ's Search Console and social routes and Otterly.AI's research tool both do.

Ask what the discovery is built from before trusting the volumes beside it.

"I need a number my finance team can audit." Buy published arithmetic. SE Visible prints the formula and Ahrefs prints the derivation. What source intelligence adds to a mention count is a separate question, and it decides what you do next.

One more filter, across all three. Ask for the provenance page before the demo, not after.

A vendor who has one sends a link in a minute. A vendor who does not sends a deck.

What This Category Cannot Do

• No tool measures the whole population. Every figure here is a sample or a model, and the vendors who say so are being accurate rather than modest.
• No tool can guarantee a mention. Engines vary answer to answer. That is why Profound runs prompts daily and SE Visible runs three per cycle.
• No volume figure is a market size. Ahrefs states this about its own metric, and it holds for every other estimate in the table.
• No provenance page proves accuracy. It proves the vendor will let you check, which is a smaller claim.

Knowing what these numbers cannot carry is what stops a directional signal turning into a forecast.

Where Klarivo Monitor Sits in This Comparison

Klarivo Monitor scores 25 out of 100 on this disclosure rubric, last in the table. That number measures what Klarivo publishes about its own method, not how well Monitor works. The product page, read on 31 August 2026, publishes four things. It runs targeted queries across five providers, each tracked independently: ChatGPT, Claude, Perplexity, Grok, and Gemini.

AEO Query Generation produces high-intent buyer queries, and reporting refreshes on a schedule the client controls. The page publishes no capture path, no volume method, and no stated limits.

Those cells are blank because nothing published fills them. Inventing a value to avoid an awkward blank is the exact failure this comparison exists to name.

All ten external tools score higher than Klarivo on this disclosure rubric. Semrush and Profound lead because they publish where their prompt data comes from in enough detail to interrogate, while Klarivo has no equivalent methodology page.

SE Visible publishes the arithmetic of its score and Klarivo publishes no formula. I scored our own platform on the same five criteria as everyone else, and this is the result.

What a disclosure score cannot capture is why a client buys Monitor. A visibility number tells you that you are absent. The citation source list tells you what the engine read instead.

Monitor reports those cited domains per answer, and closing that gap is what the wider program is for.

Want a read on your own position before you shortlist any AI visibility tool? Book a Klarivo discovery call. Fifteen minutes, and a slot you choose.

The tracking product lives at Klarivo Monitor, and how the whole program runs sets out the sequence behind it.

Frequently Asked Questions About AI Visibility Tools

Semrush, Profound, Scrunch, and Writesonic all claim real data. Which should I believe?

All four, about different things. Read what each actually claims:

• Profound names double opt-in consumer panels, so the population is defined.
• Semrush names AI search clickstream, but not which provider supplies it.
• Scrunch names third-party panel providers and restricts output to topic level.
• Writesonic names AI conversations, public forums, enterprise partner data, and industry signals, but does not identify every underlying provider.

Do AI visibility tools change anything my SEO team already tracks?

Less than it sounds. Rank tracking, crawl health, and referring domains carry over unchanged.

What changes is the query set underneath the report. A keyword list and a prompt set are not the same object, and swapping one for the other changes the number.

Can I check a vendor's provenance before I pay anything?

Yes, in about twenty minutes. Every claim about a vendor’s published methodology came from a public page.

How do I ask a vendor for its provenance and get a straight answer?

Ask four questions in writing, in this order:

• Where does the tracked prompt set come from, and who supplies it?
• Is the answer read from the consumer interface or from a provider API?
• Which numbers on this dashboard are modeled rather than observed?
• What is this metric explicitly not suitable for?

A vendor with a methodology page answers all four with links. The fourth separates them, because it costs something to answer.

My two tools disagree by twenty points. Which number goes in the board deck?

Neither on its own. Use the number whose query set you can describe in one sentence, name the method beside it, and report the second tool as a range rather than a contradiction.

Two honest samples of different populations are not a discrepancy to resolve.