How do you choose an AI visibility vendor?

You choose one by checking their claims against public documentation you can read yourself. Three things have to be true before an answer engine names a business: the engine can crawl the site, it resolves the business to one clear entity, and it sees third parties treating that business as an authority. Almost every claim a vendor makes about those three is verifiable by you, for free, before you sign. This guide is the list of checks.

Every buying guide we found ranking for this question is published by a company that sells the service. So is this one. Orion & Taurus sells this work: a 14-Day AI Answer Audit at $3,000 one time, and the AI Visibility Plan at $2,500 a month for 6 months, which doesn’t renew automatically. Research for this piece was compressed with AI, and every source here was opened and read by hand on August 24, 2026.

The argument underneath everything here is an asymmetry. Fast promises in this category cannot be falsified. Slow, specific ones can. So test every promise against what you can verify yourself, using free tools and the engines’ own documentation. A vendor worth hiring will commit to statements that could be proven wrong later.

What are you actually buying when you buy AI visibility?

Two different things are sold under the name AI visibility. One is measurement: a dashboard that reports where you stand and changes nothing. The other is work: changes made to your site, and changes to what third parties publish about you. Both are legitimate purchases, and paying a retainer for the first while believing you bought the second is the most common way money gets wasted here.

The measurement layer is cheap. Semrush sells its AI Visibility Toolkit at $99 a month per domain billed annually, tracking 25 custom prompts daily across ChatGPT, Google AI, Gemini, and Perplexity (Semrush pricing, checked August 24, 2026).

If the only artifact a retainer produces each month is a dashboard screenshot, you are paying a retainer for a $99 tool.

What has to be true before an engine will name your business?

An engine names a business when three conditions hold at once. It can reach and read the site. It can resolve the business to a single unambiguous entity, with no smear of near-matches competing for the same name. And it sees sources it does not control saying the business is good at the thing. Most proposals address the first condition, some touch the second, and the third is where the work gets slow and expensive, which is why it gets quietly dropped.

Condition one is documented by the engines. OpenAI states it in one sentence: “To make a website eligible for inclusion, allow OAI-Searchbot to crawl the site and confirm that the website host or content delivery network allows traffic from OpenAI’s published searchbot IP addresses.” (OpenAI Help Center) That is two conditions in one: robots access and network access.

Condition three has non-vendor evidence behind it. Chen, Wang, Chen, and Koudas at the University of Toronto found that for Canadian automotive queries, AI search returned 69.1% earned media, 30.9% brand-owned sources, and no social results, against Google’s 36.6% brand, 22.8% social, and 40.6% earned (arXiv:2509.08919v1, September 10, 2025). Jeremy Moser, co-founder and CEO of uSERP, puts the emphasis elsewhere, on the record: “If a GEO service does not openly tell you that success in AI visibility is 80 percent good fundamental SEO, they are selling you snake oil.” (Digiday, March 10, 2026)

Read any proposal with three highlighters, one per condition. If everything highlights the same color, you are buying one third of a program. The longer form of this rubric is on our site.

What can you check on your own site in ten minutes?

Ten minutes of checking gives you a private baseline to grade every vendor’s audit against. Open your robots.txt and look at your CDN’s bot rules. If a vendor’s diagnosis matches what you already found, they read your site. If it never mentions what you found, they ran a template.

1. Open yourdomain.com/robots.txt. Look for lines naming OAI-SearchBot, GPTBot, and ChatGPT-User (OpenAI bots documentation); ClaudeBot, Claude-User, and Claude-SearchBot (Anthropic, updated April 7, 2026); PerplexityBot and Perplexity-User (Perplexity docs); plus Bingbot, Google-Extended, and Applebot-Extended.

2. Say the distinction out loud, because it is the whole test. GPTBot and ClaudeBot are training crawlers. OAI-SearchBot, Claude-SearchBot, and PerplexityBot are the agents tied to being surfaced and linked. From OpenAI’s bots documentation: “Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links.” A vendor who treats “block GPTBot” and “block ChatGPT” as one action has the mechanics wrong.

3. Check the CDN separately. Learn Cloudflare’s three AI bot categories by name, Search, Agent, and Training, and ask which one your domain sits in. From September 15, 2026, new domains have Training and Agent blocked by default on pages that display ads (Cloudflare, July 1, 2026). A site moved onto Cloudflare in the last year may have been opted into blocking by a signup wizard.

Which promises should end the conversation?

Any guarantee of citations, rankings, or placements. The engines themselves say in writing that no such guarantee can be made, so a vendor offering one is either uninformed or is telling you something they cannot substantiate. This is the cheapest disqualifier available to you, and it costs one question.

Bing added GEO language to its webmaster guidelines in February 2026, reported this way: “GEO doesn’t guarantee citations, the guidelines note, just as SEO doesn’t guarantee rankings.” (Matt G. Southern, Search Engine Journal, February 27, 2026) Google says the same about its own results: “No one can guarantee a #1 ranking on Google. Beware of SEOs that claim to guarantee rankings, allege a ‘special relationship’ with Google, or advertise a ‘priority submit’ to Google.” (Google Search Central, “Do you need an SEO?”, updated June 5, 2026) That page also lists the questions Google suggests you ask a vendor. Send the URL to anyone pitching you and ask for written answers.

There is also no paid path into the organic answer: “Ads do not influence the answers ChatGPT gives you.” (OpenAI, January 16, 2026) A vendor selling a partnership that places you inside the answer is contradicting the engine’s own statement.

How do you tell real measurement from a screenshot?

AI answers vary from run to run, so any single answer is close to a coin flip, and any single screenshot proves nothing on its own. Real measurement is a probability computed from repeated runs over a rolling window, reported with a range. Ask two questions: how many total observations produced that percentage, and what is the confidence interval on it.

Rand Fishkin and Patrick O’Donnell, working with Gumshoe.ai across 600 volunteers and 2,961 prompt executions, found “there’s a <1 in 100 chance that ChatGPT or Google’s AI, if asked 100X, will give you the same list of brands in any two responses,” and concluded that “any tool that gives a ‘ranking position in AI’ is full of baloney” (SparkToro, January 28, 2026).

That gives you a floor to test against. Schulte, Bleeker, and Kaufmann concluded: “Practitioners should therefore use at least 7 runs per prompt per day for brand visibility monitoring, and at least 8 runs when source-level coverage matters,” aggregated over a two to four week rolling window (arXiv:2604.07585, April 8, 2026). Dmitrij Żatuchin, analyzing 12,933 responses, put brand-ranking reliability “near 0.01 for a single answer,” and found that varying the model and the language of the prompt does more for reliability than resampling one prompt (arXiv:2607.13304, July 14, 2026). That adds a question almost nobody asks: how many models, and how many phrasings of the same buyer question.

Apply the same standard to any baseline audit, including ours. A set of 25 to 40 buyer-intent prompts run once is a diagnostic. It finds fixable blockers and shows who gets cited in your place. What it cannot give you at that sample size is a statistically stable mention rate. Do the arithmetic yourself: 30 prompts run once, with an observed rate of 20%, gives a 95% interval of roughly 9.5% to 37.3%, an error bar about 28 points wide. The signal is the trend across months, and any vendor presenting a single month’s number as precision is overstating what the method can deliver.

How do you verify a vendor’s past results?

Ask for the name of the citing domain and the date it first cited the client, then open the page yourself. That is the only form of proof you can reopen tomorrow. A screenshot of a probabilistic output cannot be reproduced by anyone, including the vendor who took it.

For placements, ask for the last ten the vendor landed, by URL, then check each page for a “sponsored” or “partner content” label and for rel="sponsored" on the outbound link. Both are visible in the page source in under a minute.

Which deliverables sound technical and do almost nothing?

Four items appear on proposals with more confidence than the evidence supports: llms.txt, FAQ schema sold for rich results, generic JSON-LD sold as a route to citations, and blocking Google-Extended sold as an AI visibility change. None of them is harmful on its own. A proposal built on them is a proposal with no third-party authority work in it.

llms.txt and FAQ rich results. Google says llms.txt files “will neither harm nor help your site’s visibility” and that “Google Search itself doesn’t use them” (July 10, 2026), and it dropped FAQ rich results in 2026 (Search Engine Journal, May 10, 2026). We publish an llms.txt file and would price neither item as a deliverable.

Generic JSON-LD. Ahrefs matched 1,885 pages that added JSON-LD against 4,000 controls and measured -4.6% in AI Overviews, a significant decline, plus 2.4% in AI Mode and 2.2% in ChatGPT, both indistinguishable from zero (Linehan and Guan, Ahrefs, May 11, 2026). Google: “There’s also no special schema.org structured data that you need to add” (updated December 10, 2025). The defensible case is entity disambiguation, so an engine can tell your firm from four others with similar names. Our own work includes FAQPage and BlogPosting markup, built for that and for answer extraction, with no rich-results claim attached.

Google-Extended. Verbatim from Google: “Google-Extended does not impact a site’s inclusion in Google Search nor is it used as a ranking signal in Google Search” (July 14, 2026). Blocking it and reporting that as an AI visibility change gets the mechanics backwards.

What could a vendor do that creates risk for you?

Three off-site plays are common, and the liability for each lands on the client as well as the agency: pay-to-publish dressed as earned coverage, content produced at scale, and incentivized reviews. Each breaks a published rule, and each is checkable before it happens, so ask about all three at the proposal stage, in writing.

Pay-to-publish dressed as earned. Google’s site reputation abuse policy targets third-party pages published on a host site to trade on that host’s ranking signals, with little or no oversight from the host (Google spam policies, updated May 15, 2026). Ask whether a human editor at the publication reviewed and accepted the piece.

Bulk content. The same page reads: “Scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users.” The category prices itself per article, which is why this comes up: Red Engage sells tiers of 5, 8, and 12 articles a month (June 15, 2026).

Seeded reviews. The FTC’s final rule banning fake reviews and testimonials, 16 CFR Part 465, prohibits AI-generated fake reviews and “compensation or other incentives conditioned on the writing of consumer reviews expressing a particular sentiment” (FTC, August 14, 2024). This is federal law and it attaches to you.

How long should this actually take?

Nobody can tell you honestly, and the absence of an answer is the finding. We found no independent or peer-reviewed dataset on time-to-first-citation as of August 2026. Anyone promising citations in week one is describing something the underlying systems do not do reliably. Every published timeline we found is authored by a firm selling the service, with no sample size and no control: Goodie claims first mentions in weeks one to two (June 1, 2026); AirOps claims first mentions in 30 to 60 days (May 11, 2026).

What can be said with a mechanism behind it is that how an engine describes a business is steadier than which sources it cites. Ahrefs tracked more than 43,000 keywords and found 45.5% of citations change when AI Overviews update, while those same overviews “score 0.95 out of 1.0 for semantic consistency” (Ahrefs, November 11, 2025). Sources churn while substance holds, which makes description the more stable thing to report in month one.

A vendor who commits to a slow, specific curve has handed you something you can hold them to. A vendor promising sixty days has handed you a number with no method behind it.

Can AI visibility work hurt your Google traffic?

Yes, and the two can move in opposite directions on the same site at the same time. Ask what happens to organic search traffic while the work runs, and whether the vendor will report on it when the news is bad. The buying guides we reviewed for this piece do not raise it at all.

Lily Ray of Amsive has said case studies showing AI visibility gains often mask organic declines, and described investigating one that had lost 66% of its traffic (Digiday, March 5, 2026).

Set your expectations on what a citation is worth, too. Pew Research Center analyzed 68,879 Google searches, of which 12,593 produced an AI summary, and found users clicked a source inside the summary in “1% of all visits to pages with such a summary” (Pew Research Center, published July 22, 2025, fieldwork March 2025). Judge the return on presence in the consideration set at the moment a buyer asks.

The checks above cost you an afternoon. The wrong retainer costs considerably more, and takes longer to notice.