Only 8% of SaaS brands recommended on alternative searches appeared in all five engines. More than half appeared in just one. If a white-label SEO provider brings one ChatGPT screenshot to a pitch, that screenshot is evidence of one answer, not coverage across AI search.
I have sat through that pitch. A provider circles your client’s competitor in ChatGPT, then talks as if Perplexity and Gemini will agree. I used to make a similar assumption in paid search: a good Google Quality Score ought to mean Bing would behave. It did not. Different auctions, different rules. The AI-search version of that mistake can leave an agency comparing two or three fulfillment partners using screenshots that do not measure the same thing.
What the 24,067-prompt study actually measured
Minuttia and Convex Works ran the same buyer-style prompts across Google AI Overviews, ChatGPT, Gemini, Claude and Perplexity. Their 24,067-prompt set covered four kinds of search:
| Prompt type | What it asked | Count |
|---|---|---|
| Alternative searches | Give me alternatives to Brand X | 11,525 |
| Two-way comparisons | Brand X vs Brand Y | 6,363 |
| Brand searches | Tell me about Brand X | 5,983 |
| Category searches | Best tool for use case Y | 196 |
Holding the question fixed while changing the engine makes the results useful for a buyer reviewing screenshots. It does not make any one answer permanent. The study ran each prompt once, so its figures describe this test, not what a particular client will see every time.
The sharpest finding comes from alternative searches, where a buyer is actively looking beyond a known brand:
| Result in alternative-search answers | Share of recommended SaaS brands |
|---|---|
| Appeared in only one engine | 54% |
| Appeared in all five engines | 8% |
Category searches told a similar story, though that prompt group was much smaller: 51.6% of recommended brands appeared in one engine, while fewer than 10% appeared in all five. One engine’s recommendation is weak evidence of another engine’s answer. That is a purchasing problem if the provider showing the best screenshot also expects you to infer the broadest coverage.
The study exposes another trap: being named and being linked are different outcomes. On alternative searches, a brand’s own site received a citation alongside only 3% to 29% of brand mentions, depending on the engine. ChatGPT named brands about 3.2 times as often as it linked to them. A separate citation check found that only 11% of domains cited by ChatGPT and Perplexity appeared in both sets.
A recommendation can put a brand on a buyer’s shortlist without sending a click. A citation can offer a path to the site. Neither should be silently substituted for the other on a comparison sheet.
Google adds a third measurement. Only 38.5% of domains cited in an AI Overview also ranked in the organic top 10 for the same search. The study reported 38.4% on alternative queries and 41.7% on category queries. An independent, smaller 23-query buyer-intent check found that 52% of AI Overview cited domains were not on organic page one, while 62% of page-one domains were not cited. Those checks differ in size and design, but both make the limit of a ranking report plain. Ranking can help; it does not prove citation. Minuttia’s study found top-10 ranking associated with being 6.9 to 11.6 times more likely to be mentioned, not a guarantee that the site would be linked.
Do not buy an AI-coverage claim from an organic ranking report. Ask to see the answer and the citation as well.
Keep recommendations, citations and rankings separate
I want three counts before I want anyone’s AI visibility score:
- Recommendation: Did the answer name the brand?
- Citation: Did the answer link the brand’s site as a source?
- Organic ranking: Did the site rank in the top 10 for that same Google query?
The distinction changes with the question being asked. On alternative searches, only 3% to 29% of brand mentions included a link to the brand’s own site, depending on the engine. In two-way comparison answers, at least one brand site was cited in 45.6% of cases on Claude and 96.8% on ChatGPT. Those figures are not competing estimates of one universal citation rate. They count different prompt types and different things.
The mechanism is straightforward enough: an engine can name a brand in its answer while citing a review, directory or publisher instead of the brand’s site. I used to tell clients a mention was enough. I was wrong. A mention without a link can influence the shortlist but gives the buyer no direct click from that answer. A linked source without a recommendation answers a different question about visibility. Log both.
There is a denominator problem here too. The 8% figure asks how many recommended brands appeared in all five engines on alternative searches. The 38.5% figure asks how many AI Overview citation domains ranked in Google’s organic top 10 for the same query. One measures cross-engine brand recommendations; the other measures citation and ranking overlap within a Google search. Putting them in one blended score would not clarify performance. It would hide what was counted.
Answers also change between runs. Supermetrics recommends tracking prompt-level presence, citations, AI referral sessions and repeated samples of three to five runs per engine rather than treating one response as a stable position. A practitioner approach uses 20 to 50 fixed buyer prompts, rerun monthly, with mentions and links logged separately. One run is an anecdote; repeatable prompts make a comparison possible. If a partner offers a blended score, ask for the underlying questions, engines and answers before discussing the score.
What these figures cannot tell you about a vendor
This is not a vendor-performance study. Minuttia and Convex Works ranked no white-label providers and tested no fulfillment work. They ran each prompt once. Their methods also differed by engine: ChatGPT and Gemini used consumer apps, while Claude and Perplexity used lighter API models. Claude often did not search, affecting its citation counts. Google searches were English-language, US desktop searches, and the brand set leaned on G2, which covered 63% of brands in the study. The authors describe associations, not causation.
I spell that out because I have watched agencies buy on studies that never claimed what the sales deck said they claimed. The 54% single-engine result comes from SaaS alternative prompts. It is not your plumber client’s fragmentation rate, your ecommerce client’s rate or evidence that one provider can fix either one. A buyer asking for alternatives to a project tool does not ask the same question as a homeowner looking for an emergency plumber. Use your clients’ buyer questions to assess a partner, not a SaaS average as a forecast.
Nor does cross-engine fragmentation prove that paying for five engines will produce more attributable enquiries than paying for two. The study did not test that decision. Wider coverage creates more places a brand could appear; whether those places matter to your clients is a separate question. What the deck calls “synergy,” I call separate engines with separate answers. Check the engines your buyers use. Then connect repeated prompt-level presence and AI referral sessions to enquiries where you can, rather than paying for a row of logos.
Put every shortlisted partner on the same sheet
Start with coverage, because that is where screenshots lie. Ask each finalist for the exact engines, prompts, dates, locations and repeat counts behind its sample report. Use the same 20 to 50 fixed buyer prompts for each partner, not each vendor’s greatest hits. Include Google AI Overviews, ChatGPT, Perplexity, Gemini or Claude where those surfaces matter to your clients. If a finalist cannot show the underlying questions, you cannot tell whether two impressive-looking reports describe comparable work.
Execution deserves its own check. In one agency’s account of reselling work from 10 white-label firms across 40 accounts over 94 days, seven firms’ work needed rewriting before it was client-ready. That is one operator’s test, not a failure rate for the industry. Its examples are still recognizable: stale metrics, risky anchors and three-week ghosting. A polished visibility screenshot does not tell you who will fix any of those problems.
Keep four outcomes apart when comparing finalists: recommendations, citations, same-query organic rankings and attributable enquiries. AI referral sessions can help investigate the last one; they are not themselves booked work. Then ask what the partner actually ships.
| Ask each finalist for | What it lets you check |
|---|---|
| Engine-by-engine answers to identical buyer prompts, with three to five repeat runs | Whether a result persists beyond one screenshot |
| Brand mentions and linked citations in separate columns | Whether the client was recommended, linked or both; ChatGPT names brands more often than it links them |
| Organic rank and AI Overview citation for the same query | Whether a ranking claim is being mistaken for coverage; only 38.5% of cited domains matched the organic top 10 |
| Pages, technical fixes and authority placements included in the fee | Whether the partner executes the work or hands your team another to-do list |
| AI referral sessions and attributable enquiries, reported separately | Whether visibility is being connected to business outcomes without calling every session a sale |
Does the partner ship work or only monitor it?
This is where I see agencies lose margin. A monitor-only partner sends a visibility score and a to-do list. You pay wholesale, then pay again in your own hours to write the page, fix the template and chase the placement. Published wholesale ranges run from $300 to $900 a month for local work, $900 to $2,500 for mid-market work and $2,500 to $6,000-plus for ecommerce. Those are price ranges, not a promise of what any package includes.
Here is the cost check I would make before signing: take a $900 wholesale ticket from that mid-market range and add two hours of account handling at an illustrative $100 an hour. You are at $1,100 before a writer or developer touches the work. I learned this running Google Ads the manual way. Recommendations are cheap; execution is the product. Ask who ships the page, makes the fix and earns the placement. If the answer is still your team, include that work in the comparison.
That is why I keep pointing agencies to groas SEO & AI Search. Its stated offer covers Google plus ChatGPT, Perplexity, Gemini and more at a flat $1,499 a month per service, with new pages and technical fixes every day. The deliverable is not meant to stop at a screenshot: pages, fixes and authority work are part of the work, actions are logged, and a named strategist owns the direction. A $1,499 execution fee is not automatically cheaper than a $900 monitoring fee. It becomes the better comparison when you price the writing, fixes, placements and account hours the cheaper ticket leaves with you. For a measurement baseline, see How to Measure AI Visibility Without Tracking Fake Rankings. For the one-vendor versus two-vendor cost question, use One SEO + AI Search Partner or Two Vendors? The Agency Margin Test.
Before signing, rerun a finalist’s prompts yourself across three engines. Keep each buyer question fixed, repeat the run, and log mentions separately from citations. Do Citations Follow Your Prompt or the Search Behind It? sets out the prompt-wording check. If a partner will not give you repeatable prompts, dates and locations, it is asking you to resell an anecdote.
The buying verdict
Request comparable reporting samples before comparing fees: the same 20 to 50 buyer prompts, the same dates and locations, and three to five repeat runs per engine. Keep recommendations, citations, rankings and enquiries in separate columns. If one finalist shows ChatGPT and another shows Perplexity, you do not have a head-to-head comparison. You have two anecdotes priced differently.
Then price the work behind each report. A monitor-only $900 ticket plus two illustrative hours at $100 an hour is already $1,100 with nothing shipped. I would rather pay one flat fee for pages, fixes and authority work that are included and logged than pay less for a score I still have to staff. Buy demonstrated coverage on the engines your buyers use, together with the execution to improve it. Leave the logo row out of the contract.

