Can a free AI visibility checker tell you whether AI recommends your business?

Are free AI visibility checkers accurate?

Not as a single score, in this test. An AI agent ran five free, no-signup AI visibility checkers on 10 Chicago personal injury law firms on October 22 and 23, 2026, and only three returned results. Geoptie’s score for the same firm moved by up to 33 points from one day to the next. Loamly reported 0% for every firm while also stating that it had tested 0 prompts. Semrush and Geoptie disagreed about 2 of the 6 firms both checked. The lists of firms a checker shows are more useful than its score.

1. Repeat runs moved. Geoptie changed its score for 4 of 10 firms between October 22 and 23. Anesi Ozmon went from 0 to 33 and Malman Law from 33 to 60. Of the firm names in Geoptie’s answer lists for the same firm and engine, 39% appeared on both days.
2. A verdict without data. All 11 Loamly reports said 0% and “not being recommended.” All 11 also said “0 prompts tested” and “0 AI answers captured.”
3. One engine was always empty. Geoptie’s Google engine showed “0 brands in answer” and “No results available” in all 20 runs.
4. The banner contradicted the result. In 4 of the 7 Geoptie runs where an engine named the firm, the banner still read “You’re invisible.”
5. Checkers disagreed with each other. Semrush gave Disparti Law Group a score of 33 with 93 mentions. Geoptie gave Disparti 0 on both days, yet named Disparti 10 times in answers shown for other firms.
6. Two checkers turned the agent away. HubSpot’s form reloaded empty, and Rank Prompt showed a human-verification checkbox.
7. The research agent saw a different market. Three Claude Opus 5.5 research agents named Flanagan Injury Law Group for the best-reviews question in 2 of 3 runs. Every checker scored Flanagan 0.

None of this means the checkers are useless. AI answers are sampled, so any single check is one draw. Our synthesis of AI recommendation research found that two answers to the same question listed the same brands fewer than once in 100 runs (SparkToro, 2,961 runs), and that at least seven runs per prompt are needed before a brand’s share settles (University of St. Gallen). A free checker gives you one run. In our Reddit study of how marketers measure AI visibility, 55 of 308 threads questioned the cost or accuracy of AI visibility tools and 40 described answers that change between runs. This test shows what that looks like in practice.

Which free AI visibility checker is best?

In this test, Geoptie was the most informative free checker, because Geoptie shows the list of firms each AI engine named. Geoptie’s score moved from day to day, so run Geoptie several times and read the lists, not the score. Semrush gave a database view in about 10 seconds, limited to 3 free checks a day. Loamly was not usable here, because no prompts ran in any of its 11 reports. HubSpot’s AI Search Grader and Rank Prompt returned nothing to an AI agent.

How to use any free checker, based on these runs:

• Run it at least three times, on different days. One Geoptie run would have told Anesi Ozmon it was invisible; the next day’s run said Claude ranked it #2. Same-day re-runs do not count: Geoptie showed the same cached report when re-run 9.5 hours later.
• Read who is named instead of you. The lists show which competitors and directories the engines lean on. The score hides that.
• Check what question was asked. Geoptie does not show its prompt, and its competitor lists for Anesi Ozmon leaned to workers’ compensation firms and for McCreadyLaw to Indiana firms. A checker that asks the wrong question measures the wrong market.
• Ask a review-focused question too. The one firm of ours that the research agent recommended came up only for “best reviews.”
• Compare a live checker with a database checker. Semrush and Geoptie measure different things, so agreement between them means more than either alone.
• Make sure AI crawlers can read your site. Can AI read my site? is our free open-source checker for that.

How much do repeat runs of the same checker disagree?

Enough to change the verdict for some firms. Geoptie, run on the same 10 firms on October 22 and October 23, 2026, changed its score for 4 of them: Malman Law went from 33 to 60, Anesi Ozmon from 0 to 33, McCreadyLaw from 28 to 27, and Saks, Robinson & Rittenberg from 55 to 53. Across the three engines that returned answers, 4 of 30 engine results flipped between named and not named (13%, with a 95% interval of 5% to 30%).

The four engine-level flips were Claude for Malman Law (not named, then #9), Claude for Anesi Ozmon (not named, then #2), and ChatGPT and Claude for Saks, Robinson & Rittenberg (ChatGPT #7, then not named; Claude not named, then #14). Geoptie’s scores track how many engines named the firm: every score from 27 to 33 came from one engine, and every score from 53 to 60 from two.

The lists behind the scores moved more than the scores did. For each firm and engine we compared the two days’ lists of firm names. Of the 389 distinct names that appeared in either day’s list, 152 (39%) appeared in both. By engine: ChatGPT 21 of 86, Perplexity 55 of 123, Claude 76 of 180. Claude’s longer lists are sometimes cut off on the page with “+N more results,” so the Claude figure counts only the names shown.

Geoptie also returned one Perplexity list twice for two different firms. On October 22 the Perplexity answer for 844SeeMike and the Perplexity answer for Disparti Law Group listed the same 11 firms in the same order (844SeeMike, Disparti). That suggests both checks sent the same question, but Geoptie never shows the prompt it sent, so we cannot confirm it.

A same-day re-run of Geoptie is not a new sample. The agent ran Geoptie for all 10 firms a third time on October 23, from 22:06 to 22:12 UTC, about 9.5 hours after the second round. All ten reports were identical to the second round’s, and the nine that were timed appeared within 5 to 8 seconds, where fresh checks had taken 45 to 50 (the first, Malman Law, was read after a fixed 45-second wait and not timed). For 9 firms the SHA-256 hash of the result text matched the second round’s exactly, and for Rundio Injury Law it matched once logo placeholder initials that had not yet loaded were removed. So Geoptie serves a cached report for at least about 9.5 hours, and a re-run the same day shows the same result. The third-round runs are marked as cached in data.csv and left out of the counts and comparisons on this page. To get a new sample from Geoptie, space checks at least a day apart.

Loamly’s repeat run did not help. The agent re-ran Loamly for Malman Law on October 23. The report regenerated with progress stages (Website Analysis, Query Generation, AI Platform Queries, Results Analysis) instead of loading a cached copy, and ended in the same place: 0%, 0 prompts tested, 0 AI answers captured. Only the site-readiness score changed, from 75 to 76 (October 22, October 23).

Semrush has one run per firm so far, so it has no repeat to compare.

How much do the checkers disagree with each other?

They roughly agree on the most visible firms and often disagree on the rest. Semrush and Geoptie both found Malman Law and McCreadyLaw visible, and both found Flanagan Injury Law Group and The Horwitz Law Group invisible. They disagreed on Disparti Law Group (Semrush score 33 with 93 mentions, Geoptie 0 on both days) and on 844SeeMike (Semrush 19 with 11 mentions, Geoptie 0 on both days). Loamly said 0% for all 10 firms. An AI agent’s own research named only Flanagan Injury Law Group, a firm every checker scored 0.

Semrush allows 3 free checks a day, so it has checked 6 of the 10 firms so far: 3 on October 22 and 3 on October 23. At 06:33 UTC on October 24 its counter still showed 0 of 3 free checks left, so the daily limit does not reset at midnight UTC; the other four firms have not been checked yet. Semrush’s raw runs: Malman, 844SeeMike, Flanagan, Disparti, Horwitz, McCready.

Semrush and Geoptie measure different things

Semrush looks up a domain in its database of more than 26 million prompts and responses. Most of the mentions Semrush counted came from Google’s AI Overviews: 159 of the 280 mentions across the four firms with any (57%), and AI Overviews were the largest single platform for three of those four. For all four, the top “best performing topic” names the firm or a person behind it: “Steve Malman Family and Legal Profile,” “Mike Agruss Law and Legal Services” (for 844SeeMike), “Disparti Law Group Legal Services and Reviews” and “Mccready Law and Injury Attorneys.” So part of a Semrush score reflects answers to questions that already name the firm. Semrush’s most-cited pages were often blog posts far from hiring a lawyer, such as a post on the stages of sepsis for Malman Law and one on when a child can sit in the front seat for 844SeeMike.

Geoptie asks AI engines a category question that Geoptie infers from the website, and shows who was named. Whether a firm counts as visible then depends on which question Geoptie picked. Anesi Ozmon’s runs leaned to workers’ compensation firms. McCreadyLaw’s leaned to Indiana firms, and Semrush’s topics for McCreadyLaw include Chesterton, Indiana, too.

Disparti Law Group was named 10 times, in 9 of the runs Geoptie made for other firms, and never in its own two runs. Malman Law was named 13 times elsewhere and 3 times in its own. A firm’s own check is only as good as the question the checker chose for it.

Names can also collide. For The Horwitz Law Group, the answers were full of a different firm, Horwitz, Horwitz & Associates: on October 22 all three engines named Horwitz, Horwitz & Associates, on October 23 Perplexity did, and no engine named The Horwitz Law Group on either day. Geoptie correctly scored The Horwitz Law Group 0, but a quick reader could mistake one firm for the other. Horwitz, Horwitz & Associates was the most-named firm across all Geoptie answers: 34 times, in 18 of 20 runs. In the research agent’s 15 answers it appeared twice.

Semrush also disagreed with itself once. For The Horwitz Law Group’s domain, Semrush’s headline said 1 citation, while its own Top Cited Pages table listed pages with 2, 1 and 1 citations (raw run). The page also shows trend and score panels inside upsell blocks with the same figures for every firm; we did not count them.

Why did some checkers not work for an AI agent?

Two of the five checkers the agent tried gave the agent no result. HubSpot’s AI Search Grader reloaded with an empty form both times the agent submitted it; the page loads a bot-detection script from CHEQ, but the cause is unverified. Rank Prompt showed a Cloudflare “Verify you are human” checkbox, and the agent does not tick human-verification boxes. Neither was retried or worked around. For anyone who plans to have an AI agent run these checks, that is a finding in itself.

• HubSpot AI Search Grader, October 22, 2026, 08:23 UTC, for Malman Law (company name “Malman Law”, location “Chicago, Illinois”, products or services “Personal injury law”, industry “Legal services”). The agent submitted the form once by clicking and once by pressing Enter after declining cookies. Both times the page reloaded with the inputs in the address and an empty form. HubSpot’s terms allow automated access at a human rate: they bar systems that send more requests “than a human can reasonably produce in the same period by using a conventional on-line web browser” (§2(v), terms last updated April 14, 2026, checked October 21, 2026). Run notes, screenshot.
• Rank Prompt, October 22, 2026, 08:25 UTC, for Malman Law. A Cloudflare Turnstile checkbox appeared under the form and the results stayed on “Asking the engines.” Rank Prompt’s terms list as a prohibited use: “Use automated tools to access the Service in a way that exceeds reasonable use” (§8, terms last updated October 20, 2026, checked October 21, 2026). Screenshot.

Our crawler study found the same pattern on the firms’ own websites: 13 of 78 sites refused a request carrying the name GPTBot, ClaudeBot or PerplexityBot in all five runs.

Which checkers did we leave out and why?

Six. Ahrefs and Mangools ask users in their terms not to use their sites by automated means except through their APIs, so the agent did not run them; a person may run both by hand later, and this page will be updated if so. The terms of Arobis and GeoBuddy also prohibit automated use. Honeyb shows a fixed snapshot captured on October 2, 2026, so repeat runs cannot differ. Howseen’s terms say its free report was discontinued on October 10, 2026.

Terms of service are what each vendor asks of its users. We followed them.

What did an AI agent’s own research find?

Three Claude Opus 5.5 agents researched five buyer questions on the live web on October 23, 2026, without being told which firms we tested, and named 33 different firms in 93 mentions. Only one of our 10 firms appeared: Flanagan Injury Law Group, named in 2 of 3 runs for “Which Chicago personal injury law firms have the best reviews?” with a Google rating of 4.9 from about 133 reviews, taken from Expertise.com. This is one AI agent’s research, not what ChatGPT, Perplexity, Gemini or the Claude app would tell a consumer.

Each agent ran at high effort in a fresh context with only web search and web fetch, and was told to answer as a careful assistant helping a consumer in Chicago, naming 3 to 10 firms per question. The three were launched together at about 19:47 UTC. The instructions, word for word, are in prompt.md, and the answers, searches and sources in run1.json, run2.json and run3.json.

The most-named firms across the 15 answers were Corboy & Demetrio (10), Clifford Law Offices (9) and Power Rogers (7), then Abels & Annes and Romanucci & Blandin (6 each). For the questions about the best lawyers and the car crash, the same four firms came back in all three runs. For the best-reviews question, no firm came back in all three, and 19 different firms were named.

That review question is where Flanagan Injury Law Group appeared, ranked #3 in run 1 and #6 in run 3. Both runs cited Expertise.com’s Chicago listing for the rating. Flanagan scored 0 in Geoptie on both days, 0 in Semrush and 0% in Loamly. A firm can be missing from a checker’s category question and still surface when the question is about reviews.

The research leaders also differ from Geoptie’s. Horwitz, Horwitz & Associates and Briskman Briskman & Greenberg each appeared in 18 of 20 Geoptie runs. In the agent’s 15 answers they appeared twice and once.

How was the test run?

An AI agent (Claude, in Claude’s built-in browser, inside a Claude Code session) ran every check one at a time at human pace on October 22, 2026 (08:23 to 08:54 UTC) and October 23, 2026 (12:35 to 12:51 UTC, plus a third Geoptie round from 22:06 to 22:12 UTC). The agent used no AI APIs and no paid keys, declined optional cookies, entered each firm’s website or name exactly as listed below, and saved the page text of every result and a screenshot of each failure. The 10 firms are every 10th row of the 100 firms in our crawler study.

1. Order and pace. On October 22 the agent tried HubSpot, Rank Prompt, Semrush, Geoptie and Loamly for Malman Law, then ran Geoptie for the other nine firms, Semrush for two more and Loamly for the rest, one check at a time. On October 23 the agent ran Semrush for three more firms, Geoptie for all 10 again and Loamly once more for Malman Law, and from 22:06 to 22:12 UTC, about 9.5 hours after the second round, Geoptie for all 10 a third time. In the first two rounds, Geoptie checks started about 70 seconds apart. Start times are in runs/agent/log.csv.
2. Saving. For each run the agent saved the result text word for word in a .txt file, and a screenshot for every Geoptie and Semrush run and for the failures. Loamly reports were re-read about 2 to 14 minutes after submission to catch the final state.
3. Terms followed. Geoptie’s terms ban only “automated access that circumvents plan allowances” (§5, terms last updated August 23, 2026). Semrush’s terms set a human-rate standard for automated systems (restriction (k), terms last updated August 25, 2026). Loamly’s terms (December 2024) have no clause on automated use. These three were checked October 21, 2026. HubSpot and Rank Prompt are quoted above.
4. Counting. method/analyze.py parses the saved runs into data.csv (one row per run) and mentions.csv (one row per firm named in a Geoptie answer or a research answer: 541 Geoptie rows and 93 research rows). A firm is matched to our list by name fragments (for example “malman”, “disparti”, “844seemike” or “agruss”). Names in Geoptie lists cut off with “+N more results” are not counted. Geoptie has 30 rows in data.csv, one per run; the 10 third-round runs, served from cache, are marked yes in its cached column and left out of mentions.csv and of the counts and comparisons on this page.
5. Comparing names. For the list overlap and the agent-research agreement, names were lowercased and stripped of a leading “The,” punctuation and LLP, LLC, P.C., PC or Ltd. Variant names such as “Katz Friedman” and the full “Katz, Friedman, Eisenstein, Johnson, Bareck & Bertuca” still count as different, so overlap is slightly undercounted.
6. The interval. The 95% interval for 4 flips in 30 engine results is a Wilson score interval. The 30 are three engines (ChatGPT, Perplexity, Claude) times 10 firms; Google is left out because it never returned a list.

p = k/n = 4/30 = 0.133; z = 1.96
interval = [p + z²/2n ± z·√(p(1−p)/n + z²/4n²)] ÷ (1 + z²/n)
= 0.053 to 0.297, about 5% to 30%

The interval treats the 30 engine results as independent draws, which they are not quite: they come from 10 firms and one checker. Read it as a rough guide to how unsure 30 observations leave us.

What about open-source AI visibility trackers?

We read the setup instructions of three open-source trackers, Elmo, OneGlanse and Gego, on October 21, 2026, but did not install them. All three need paid AI API keys or accounts with the AI products before they return a first result, and the agent does not enter real API keys or sign in to accounts. Elmo and OneGlanse are MIT-licensed; Gego is GPL-3.0.

OneGlanse and Elmo’s scraping providers drive the consumer chat sites, whose terms may restrict automated access; that is a risk for each user to weigh. Answers from a model API can also differ from what the consumer app shows.

Sources

Checker pages were used on the dates of the runs. Terms pages were read on the dates given.

• Geoptie (Textintel FZE), Free AI Visibility Checker, used October 22 and 23, 2026. geoptie.com/ai-visibility-checker. Terms of service, last updated August 23, 2026, read October 21, 2026. geoptie.com/terms-of-service
• Semrush Inc., AI Search Visibility Checker, used October 22 and 23, 2026. semrush.com. Terms of service, last updated August 25, 2026, read October 21, 2026. semrush.com/company/legal/terms-of-service
• Loamly, Inc., AI visibility check, used October 22 and 23, 2026. loamly.ai/check. Terms, last updated December 2024, read October 21, 2026. loamly.ai/legal/terms
• HubSpot, AI Search Grader, used October 22, 2026. hubspot.com/ai-search-grader. Website terms of use, last updated April 14, 2026, read October 21, 2026. legal.hubspot.com/website-terms-of-use
• Rank Prompt LLC, Free AI visibility checker, used October 22, 2026. rankprompt.com. Terms, last updated October 20, 2026, read October 21, 2026. rankprompt.com/terms
• Ahrefs Pte. Ltd., Free AI visibility checker. ahrefs.com/ai-visibility-checker. Terms, last modified September 14, 2026, read October 21, 2026. ahrefs.com/terms
• Mangools, AI Search Grader. mangools.com/ai-search-grader. Terms and conditions, last updated June 20, 2023, read October 21, 2026. mangools.com/conditions
• Arobis AI, ChatGPT visibility checker. arobis.ai/chatgpt-visibility-checker. Terms of use, last updated July 2026, read October 23, 2026. arobis.ai/terms-of-use
• GeoBuddy, check. geobuddy.co/check. Terms, last updated February 7, 2026, read October 23, 2026. geobuddy.co/terms
• SIA Honeyb, ChatGPT visibility checker. honeyb.ai/tools/chatgpt-visibility-checker. Terms of service, last updated March 4, 2026. honeyb.ai/terms-of-service
• Howseen AI, AI visibility checker. howseen.ai/tools/ai-visibility-checker. Terms, last updated October 15, 2026, read October 23, 2026. howseen.ai/terms
• Expertise.com, Chicago personal injury lawyers listing, as cited by the research agent on October 23, 2026. expertise.com
• Elmo, OneGlanse and Gego READMEs, read October 21, 2026. github.com/elmohq/elmo, github.com/oneglanse/oneglanse, github.com/AI2HU/gego
• Copy for Originals, “Do law firm websites block AI crawlers?” (the source of the 10 firms), October 14, 2026. tools.copyfororiginals.com/law-firm-websites-ai-crawlers-study/
• Copy for Originals, “What decides whether an AI recommends you, and what to do about it,” October 17, 2026. tools.copyfororiginals.com/what-decides-ai-recommendations/
• Copy for Originals, “Useful diagnostic, terrible KPI”: what 200,753 Reddit posts say about measuring AI visibility, October 16, 2026. tools.copyfororiginals.com/ai-visibility-measurement-reddit-study/

Raw files: data.csv, mentions.csv, firms.csv, method/checkers.csv (the 11 checkers, their limits and terms), method/terms-quotes.md, method/open-source-trackers.md, runs/agent/log.csv and the page text and screenshot of every run, linked from the tables above.

How to cite: Copy for Originals, “Can a free AI visibility checker tell you whether AI recommends your business?”, October 24, 2026, https://tools.copyfororiginals.com/free-ai-visibility-checkers-tested/. Data licensed CC BY 4.0.