Why are there no reviews for AI visibility agencies?

Reviews are scarce for four reasons that have nothing to do with the quality of the work, and knowing them changes what you should go looking for instead.

The first reason is age. Answer engine optimisation became a service line after ChatGPT search launched in late 2024, so most providers have a client roster measured in months rather than years, and a review profile needs a long tail of finished engagements to mean anything.

The second is where the reviews would have to live. G2 is organised around software products, and a service engagement has no product page to attach a star rating to. Clutch does cover agencies and verifies that projects happened, which makes it the better of the two for this category, and it is still thin on firms that sell AI visibility specifically.

The third is confidentiality. AI visibility work touches a company's positioning, its pricing page and sometimes its competitive weaknesses, and buyers in that position often will not be named. We are in exactly that position ourselves: our early results page carries two anonymised case studies rather than named clients, because the engagements are covered by non-disclosure agreements and none of those clients has agreed to be a reference in public.

The fourth is what fills the gap. A search for buying queries in this category on 9 November 2026 returns a first page made up almost entirely of guides and rankings published by firms that sell the service, several of them posts about the red flags to watch for in firms that sell the service. None of that is dishonest in itself, and all of it is marketing. Treat an agency-published ranking as a source of names and nothing more.

So vetting here cannot be done by reading what other buyers said, because other buyers have not said anything yet. It has to be done by inspecting the vendor's method, its reporting and its contract, all of which are available to you before you sign.

Can you trust a "best AEO agencies" list when an agency published it?

An agency-published list is useful for assembling a shortlist and useless as a ranking, and there is now a measurement that explains why the people publishing them may be getting less out of them than they think.

Lily Ray ran 100 B2B "best [category] software" queries through Google AI Overviews on 15 April, 15 May and 8 June 2026, using Ahrefs Brand Radar to capture the answer text and the cited sources. Where a brand's own self-promotional listicle was cited as a source, that brand was left out of the actual recommendation in 224 of the 323 cited self-promotional listicles, which is 69%, according to her study published on 17 June 2026. The competitors named inside those lists were recommended instead. Her own reading is that the recommendation decision is anchored to how much of the web is talking about you, rather than to what your page claims.

Read that as a buyer and the implication is direct. The list in front of you was written by a firm whose position in it is the one thing the author controlled, and the engine that showed it to you may well be recommending somebody else named on the same page.

Four things make such a list worth reading anyway. Criteria stated before the options, so you can see what was being judged. A line per option saying who it suits, which is the part buyers actually self-select on. A visible publication date, because prices and feature sets in this category go stale in months. And a plain disclosure of who published the page.

We publish one of these lists and we are on it. Our comparison of AEO agencies includes SIGNALS, states that we publish the page, and carries a section on when to choose a different kind of provider. Read our own row there the way you would read any vendor's account of itself.

What should you ask a vendor instead of asking for reviews?

Ask a vendor for the mechanics of the measurement and the mechanics of the exit, because both are verifiable in a single call and neither depends on anyone's opinion of the firm.

Mike King of iPullRank published a list on 4 November 2026 of what to make a vendor show you before signing, and it is the most directly usable version we have seen. His four asks are "Which engines it covers, how many prompts it samples, and how often it collects data", "The path from a specific finding to a proposed action, with the evidence behind it", "The review your team must do before any automated fix goes live", and "What you can export, how much history you keep, and what happens when the subscription ends". King runs an agency in this category, which is worth holding in mind when reading his conclusion that a platform alone is not enough, and his question list stands on its own.

Two more questions are worth adding, and both come from how this work actually goes.

Ask to see a sample report with the raw answers in it, not only the score. A score is a summary of someone's judgement about your pages; the engine's actual answer, with its cited sources, is the evidence. A vendor who will not show you the answers is asking you to trust the summary.

Then ask what happens when the honest finding is that nothing on your pages is the problem. A page can be well built, correctly crawled and still not cited, because the engine never retrieved it or because the model does not know your brand. A vendor who always comes back with page fixes is not reading the page, and a vendor who will say "this one is not a content problem" is telling you something about how they work. Our own scoring abstains rather than inventing a fix when a page is too thin to ground one, which is the same discipline seen from the supply side.

Does it prove anything if the agency is cited by AI itself?

Being cited by AI proves very little about an agency, and the advice to run that test, which appears in most guides on this subject, is weaker than it sounds for two measured reasons.

The first is that a single check is noise. SparkToro ran repeated brand-recommendation prompts with Gumshoe.ai, using roughly 600 volunteers across five survey rounds and running individual prompts about 100 times each. On Google AI Mode the chance of getting the same brand list twice in 100 tries was under 1%, and it took roughly 124 queries before the same two brands appeared together twice, according to the SparkToro session published on 8 May 2026. Average visibility for the top three brands ran about 64% on ChatGPT, 73% on Claude and 68% on Google AI, with a standard deviation near 20 points. Gumshoe.ai sells AI tracking software, so the finding that presence needs many runs to measure is one that suits its business, and the design is still the best-powered version of this test we have read.

The second reason is that citation is not a competence signal. Mahe Chen and Nick Koudas studied 158,099 generative engine answers carrying 468,037 citations and found that which pages the engine retrieves explains about 32% of the variation in citations while page properties explain 2.8% (arXiv 2610.15033, submitted 10 October 2026). They also report that ChatGPT writes brands it already knows into its own searches and names brands with no supporting page at all. An agency that gets named may be older and better known rather than better.

Our own category research shows how badly one engine can mislead. Across 27 buyer queries in three pharmaceutical skid categories, one fabricator was named in the top three on Google, Perplexity or Claude in twelve of fourteen queries in its lane and scored zero on ChatGPT in eleven of those same fourteen. A buyer who checked one engine once would have reached the opposite conclusion from a buyer who checked another.

So ask the question differently. Rather than testing whether the agency appears, ask it to show you its own prompt set, its own numbers across all four engines, and the interval around them.

What does an honest answer about results look like?

An honest results answer separates what the vendor moved from what it hopes will follow, and says plainly which of the two the number describes.

Here is the distinction applied to our own work, which is the only engagement we can show you in full. Our early results page reports one client's SIGNALS audit score moving from 17 to 68 and a second more than doubling, and then says in its own words that the audit score is an input metric that tells you how well a site is structured for citation, while share of voice is the outcome metric, and that the outcome tracking was still running. A strong audit score stacks the odds and does not guarantee a citation.

Most results claims in this category blur those two. A chart of score improvements presented as evidence of citations is the common version, and a chart of citations presented without an interval is the other. Both are worth asking about directly: which of these numbers is something you changed, and which is something you observed?

Listen for three more things. Whether a reported citation figure comes from one reading or many, since a single reading cannot be distinguished from sampling. Whether the vendor will say what did not work, because an engagement with no failed experiments in it has not been described to you accurately. And whether the forecast has a shape: page work moves in weeks, third-party mentions and brand familiarity move over months, and a proposal that promises both on the same timeline has not been thought through.

Guarantees are the clearest tell. No provider controls what an engine retrieves, so a guaranteed citation is a promise about somebody else's system. Our guide on whether an AEO agency should guarantee results covers what can honestly be committed to instead, which is process, cadence and reporting rather than outcomes.

What evidence can you actually check before you sign?

Six kinds of evidence are available to a buyer in this category, and each proves something narrower than it appears to.

Evidence	What it can prove	What it cannot prove
Named client case study	That a real engagement happened and someone will say so	That the result generalises; rare in this category because of non-disclosure agreements
Anonymised case study	What the vendor measures and how it reports movement	That the client exists, unless a reference call is offered
An agency-published "best of" list	Which firms exist, and what criteria the author thinks matter	Any ranking; the author chose its own position
The vendor's own AI visibility	Little on its own, and something if shown across engines with repeated runs	Competence, since brand familiarity drives naming independently of quality
A sample report with raw engine answers	How the vendor reasons from an answer to a recommended change	That your own category behaves like the sample
Contract, export and exit terms	Exactly what you keep: prompt sets, history, page changes	Nothing, and this is the most checkable item on the list

Work down that table in order of how much it is worth to you rather than in order of how impressive it looks. The last row is the one most buyers skip and the only one that is fully verifiable in writing before any money changes hands.

One practical note on references. Ask for a reference whose engagement ended, not only a current client, and ask them what the vendor said when a month went badly. A happy current client tells you about onboarding; a finished engagement tells you about the work.

When should you choose a different kind of provider?

Choose someone else in four situations, and all four are easier to spot before you start than six months in.

If your team already has the expertise and the hours to investigate findings and ship page changes, buy software and keep the budget. King's own framing is worth quoting: "A platform-only approach works when your team already has that expertise and capacity" (iPullRank, 4 November 2026). Our comparison of doing AEO in-house against a tool or an agency walks through where each model breaks.

If your procurement process needs a large vendor with a formal security review, a long trading history and a bench of people, a boutique firm will fail that process no matter how good the method is. That is a real constraint and not a judgement about either side.

If the actual problem is that almost nothing on the web mentions your company, the first lever is coverage rather than pages, and a public relations or analyst relations firm will move it faster. Our guide on third-party mentions sets out how to tell which side your problem sits on.

And if the engines cannot fetch or read your site at all, the first invoice should go to a developer. A page that cannot be retrieved is never assessed on anything else, and no amount of content work changes that.

Where none of those four fit, the thing to buy first is a diagnosis rather than a retainer. A free visibility assessment runs your buying questions across ChatGPT, Perplexity, Google AI Overviews and Claude and shows which phrasings your competitors already own, which is enough to tell whether the problem is your pages, your presence or your category.

What else do people ask about vetting an AI visibility agency?

Buyers ask these five questions most often once they start looking for evidence about a provider in this category and cannot find any.