Blog · 24 September 2026 · By HeardOf

AI visibility tools: what nineteen of them say about how they get their answers

How do AI visibility tools collect their answers?

Thirteen of nineteen say, on their own pages as read on 24 September 2026, how they collect their data: eight through the chat interface, four through both an API and the interface, and one through an API. Six say nothing on the pages we read.

The nineteen are the set we track, and their prices are in what nineteen AI-visibility tools charge; this post is the method column. Seven questions were put to each vendor's own docs, help centre, knowledge base, API reference, status page, changelog and product pages; they are the columns of the two tables below. A negative means not stated on the pages we read, which are listed by URL, time and status in the research file behind this post; the rule for what counts as a statement is at the end. What each vendor's pages say the product shows you — the answer text, the brands you did not add, the URLs each answer cited — is in can you see the answer, and who got named instead, read the same day.

The interface camp is the larger one. Profound: "Other tools pull AI responses from the API. We capture directly from the browser." Peec: "Peec AI uses UI scraping technology to interact with AI platforms exactly as real users do", and, for most engines, "the platform decides which model to use and whether it runs a web search." Otterly, in a help article dated 17 July 2026: "posting questions and prompts through the website of the AI search (like chatgpt.com etc)", with a footnote that Claude alone is read through an API. Semrush: "Prompt responses are captured from real requests and not via any APIs of LLMs." Ahrefs' help centre: it "enters those questions into the web version of each AI chatbot using the default model", Claude excepted, and its methodology post says the same. Writesonic, Surfer and SE Ranking use the word scrape or the phrase graphical user interface.

Four say they use both an API and the interface, and two of them say which engine gets which. Rankscale's pricing calculator labels "GUI engines AI Overview GUI, Grok GUI, Copilot GUI" against "API engines P Sonar, P Sonar-Pro, P Sonar-R-Pro, GPT-5, Gemini 3.0F, Gemini 3.1P, Mistral Large". Athena's API changelog, in an entry under August 2026, says Meta AI "is reached through Meta's Model API rather than scraped from a country-local session" and that the same is true of Claude and DeepSeek, which puts its other engines on the scraped side. Evertune tracks "through the model's API and its consumer app", both routes on every model. Scrunch lists "browser automation and official platform APIs" without saying which engine gets which. One says API: Conductor's docs say "We use official APIs provided by the AI search engines we support to get the content from responses", and its product page calls that "an official API-first approach for data collection wherever possible".

Which tools name the model they query?

Six of nineteen name a version string somewhere on their own pages; thirteen do not. Naming one is not the same as naming all: Ahrefs names Claude Sonnet 4.6 and calls everything else "the default model", and Peec names versions only for the paid API models, since for the scraped engines "the platform decides which model to use".

Two are not counted, for stated reasons. Semrush says its toolkit "analyzes responses generated by the latest model of ChatGPT in search mode", a mode and a product but not a version; Surfer says it does not pick a model at all, on the grounds that its customers do not either. Neither is a version string, which is the thing that tells a buyer which model produced last Tuesday's number.

How many times a day does each tool ask a prompt?

Two say once a day in so many words, Profound and Peec. Three use wording or an example that amounts to one run per prompt per engine per day: Otterly, whose KPI definitions count "each execution of a day/prompt/service" as one; AmICited, whose pricing page counts "one prompt run on one engine, once"; and SE Visible, which "checks all 200 prompts in all 5 LLMs every day". Five give a schedule: Scrunch's default (daily for fourteen days, then every 72 hours, refreshable on demand), and customer-chosen schedules at Rankscale, Ahrefs, Athena and Conductor — Rankscale from hourly to monthly, Ahrefs daily, weekly or monthly, Athena weekly or monthly, Conductor daily, weekly or monthly per topic. Evertune gives a sample count with no cadence: "Evertune samples every prompt 100 times per model", set against "most GEO tools" that "sample each prompt once a day", and that comparison is the only place cadence appears on the pages we read. Surfer says "multiple queries per day per model, then average the results" and gives no number. Five say daily with no count — Goodie, Semrush, HubSpot, Writesonic and Similarweb — and two say nothing on the pages we read, Bluefish and BrightEdge.

Why the number matters is in our post on which engines we check, as updated 23 September 2026, which cites a published study by Schulte, Bleeker and Kaufmann (arXiv, 8 April 2026) putting the standard error of a single run at 0.370 and recommending at least seven runs per prompt per day. One run a day is what we do too, per that post; its claim that one answer per prompt estimates a rate without bias is a claim about our own system, not about any vendor in the table. A vendor that states no cadence leaves a buyer unable to tell what its dashboard is a sample of.

What happens when Google shows no AI Overview, and when a fetch fails?

One of nineteen states a rule that names AI Overviews and says how the prompt counts, and none states how a failed fetch counts. The one is Otterly, whose KPI definitions, dated 4 September 2026, say: "Special case: Google AIO When AIO are not triggered, they do not count towards that formula." A prompt without an overview leaves Otterly's denominator. Conductor's docs define the case as a status: "No AI response generated" means "Conductor successfully submits a tracked prompt to an AI search engine, but no AI response is returned. This is most common for Google AI Overviews, where AI responses are generated only for 50–70% of prompts", without saying how such a prompt enters coverage. Profound's metrics documentation states a general rule, "Runs in which no brand was mentioned are excluded from both the numerator and the denominator", without naming AI Overviews; if Profound records a prompt with no overview as a run, that run falls out of its visibility score along with any other answer that names nobody, and whether it does is not on the page.

The failed-fetch column carries no rule because no vendor's page we read states one: whether a fetch that fails is dropped, retried, counted as an answer with nobody in it, or left as a gap, and each of those moves the rate a customer sees. Conductor's docs define "No AI response collected: Conductor could not successfully collect a response from the AI search engine for the tracked prompt" as a status distinct from an empty answer, with no sentence on how either enters coverage. Athena's API reference defines its denominators as "successfully analyzed responses", Ahrefs notes it is "temporarily unable to gather new data" from Grok, HubSpot's Grader page mentions "retry logic"; none is a rule for the count. Ours is in the post linked above, as updated 23 September 2026: a failed call is retried, a prompt whose retries all fail is a failure for that engine and not an empty answer, and enough failures take the engine out of the headline.

Has any vendor said anything about Google's scraper block since 13 September?

Not on any page we read, the help centres, status pages and changelogs listed in the research file included. From around 13 September 2026 Google may have become more successful at blocking scrapers and tracking tools, and vendors that collect Google results pages reported problems; the reporting is linked in our post on which engines we check, which said that as of 19 September 2026 it had seen no statement from any of the fifteen vendors that list AI Overviews on what the block did to their collection. Five days later there is nothing dated on any page we read.

Two vendors did publish something dated after 13 September about a Google AI surface, and both concern citation counts, not collection. Profound's status page, read on 24 September, marks its Google AI Mode component "Google AI Mode citations are affected", with a degraded entry on every day from 2 September, when the incident opened, to 24 September, the day we read it, and an update on 18 September: "We've made progress in understanding the variability affecting Google AI Mode citation data and are validating an updated approach to improve data stability." Its Google AI Overviews component shows no degraded day in the same period. Evertune's post of 23 September says it "tracked roughly 180,000 AI Mode responses between August 11 and September 3"; Google blamed the zero-citation answers on "a bug in the new Gemini 3.8 Flash model", Evertune says the decline began three weeks earlier and that "its timing matches the /goto rollout", the redirect links Google began wrapping around results in late August, and it reports that AI Overviews' sources per response "declined around 15%" over the same period. Neither mentions scraping or blocking, and we are not claiming either is about the block; a reader who wants to know what happened to AI Overviews data in mid-September will find nothing dated on any page we read.

Do the vendors' own pages agree on how many engines they check?

Six of nineteen give a number on one page that another page, or the same page, contradicts, with a different number or with a named list of a different length; three more give no number and a different list on each page; ten are consistent or give a single list.

The six: Otterly says "six AI platforms" in its home-page FAQ and "six major AI search engines" in a help article dated 7 July 2026, and "seven AI engines" on its features page, while a help article dated 9 September names seven. Goodie's pricing page says "Up to 13 engines" in one tier and "all 12 engines" in the FAQ below it. Rankscale says "17+ AI engines" on three pages, while its pricing calculator lists six engines and "+ 10 more engines", and the ten it names make sixteen. AmICited says "all 8 engines" on its rank-tracker and pricing pages, while its home page lists eleven in its FAQ and twelve in its coverage table, seven daily and five weekly. Peec's pricing page says "Up to 13 LLM models tracked" on its Enterprise card and names fourteen models in the Enterprise column of the table beneath it and gives its active models as unlimited, as served on 24 September 2026; the six self-serve models in that table include Naver AI and not Perplexity, where its ai-instructions page's six default models include Perplexity, a difference of membership the rule does not count. The three without a number are Semrush, whose knowledge base names four platforms behind its prompt database in two articles, five and then six in a third, and whose solutions FAQ names six; Conductor, whose docs list nine engines while its product pages list three to five; and BrightEdge, with a different list on each of three pages. Profound's pricing page says "Up to 9 Answer Engines tracked" and names nine, while its status page carries a component for eight, without Microsoft Copilot; its features page names nine with Grok and without Exa Search, a difference of membership the rule does not count.

None of this is a finding about what the tools do; two pages written at different times can both be honest and still disagree. It is a finding about what a buyer can check: on six sites, the count is a question the site answers two ways.

Correction, 24 September 2026. This post went out counting Peec as consistent, from the lists on its ai-instructions page; its pricing page, read the same day, says up to thirteen on the Enterprise card, while the table beneath names fourteen for Enterprise and gives its active models as unlimited, so Peec is the fifth and the consistent count is eleven, not twelve. Later the same day, Profound's status page, which lists a component for eight engines against the pricing page's nine, made it the sixth, and the consistent count ten; the same read moved the Google AI Mode entry from 1–23 to 2–24 September, the dates the status page gives.

What counts as a stated methodology, and how was this read?

A statement counts when it is on the vendor's own domain and answers the question for the product as sold, and the read was one pass on 24 September 2026, between 13:54 and 15:06 UTC, with curl and a browser user-agent, no JavaScript executed: 1,690 requests to the nineteen vendors' domains and four to our own posts, each listed with its URL, time and HTTP status in the research file. The domain rule admits a product page, docs, a help centre, a knowledge base, an API reference, a status page, a changelog or a methodology page, and a blog post only when its subject is the product's own method, which is why Ahrefs' methodology post is in and Surfer's study of scraped-versus-API answers, updated 17 September, is not. Testimonials, figures inside product mock-ups, third-party reviews, what a vendor says about a competitor, and a sentence that answers a different question do not count. A model version means a version string, so "the latest model" and "Sonnet" do not. A rule for a missing overview or a failed fetch means a sentence about how it enters the figures, so a study of how often overviews appear does not, and a status label that names the case without the count is recorded as that. A dated statement about the block means one dated on or after 13 September 2026 whose subject is Google blocking or changing what tracking tools can collect. An engine-count disagreement means a number on one page contradicted by another number, or by a named list of a different length, for the same product.

The read covered the docs, help-centre, knowledge-base and changelog pages listed in the research file, reached from each vendor's llms.txt, docs index or help index, plus every sitemap we could read, filtered for pages dated on or after 13 September. Sites that answered with a bot challenge were left there — SE Ranking's help centre, Similarweb's support centre and Athena's status page answered 403; Goodie's help subdomain and Rankscale's docs subdomain do not resolve; the HubSpot knowledge-base addresses we tried answered 404 and its search page served no article links; 45 of Writesonic's docs pages answered 429 on three attempts — which is why every negative in the tables is scoped to the pages read. At least four vendors carry text addressed to AI readers — Peec's ai-instructions page, AmICited's file for language models, Otterly's footer link inviting AI to learn about it, and the banner on every Writesonic docs page — and all were read as data.

Otterly's and Semrush's cells were re-read on 24 September 2026 rather than copied from HeardOf vs Otterly and HeardOf vs Semrush; the wording matched, and so did the two Semrush database figures the Semrush post recorded: 317 million in its knowledge base and 239 million on its Semrush One page, both there on 24 September.

Everything this post says about our own system comes from our own post on which engines we check, as updated 23 September 2026, and carries that provenance in the sentence that says it; you cannot check it. Everything else carries a vendor's page and the day it was read, and the research file carries the time.

Can you see the answer, and who got named instead? What nineteen AI visibility tools' pages say