What ChatGPT Recommends When You Ask for an SEO Agency

Last updated 7 September 2026 with nine additional weekly samples, the review-panel finding, and a correction to our earlier conclusion.

Across fourteen weekly samples of the same buyer question, ChatGPT never once built its shortlist primarily from the agencies’ own marketing content — it built it from third-party directories, review platforms, forum threads, and, in roughly a third of runs, agencies’ own service pages read as reference documents rather than as sales pages. The shortlist also changed substantially week to week, with only a minority of names persisting between runs.

We ran this test because we needed the answer for our own business, and because the standard advice — publish more content and you will get cited — did not match what we were seeing. Aggarwal et al. (Princeton, 2023) established that page-level changes like adding citations and statistics can lift visibility inside generated answers by up to 40%. What that research does not tell you is which pages the model reaches for in the first place. That is what we set out to log.

How the test was run

Every week since early June we have asked ChatGPT, with web search enabled, the same question a real buyer would type: a business in Provo, Utah wants an agency that can handle Google rankings and get them mentioned by AI assistants — who should they shortlist? We recorded every agency named, every source URL cited, the order of the recommendations, and — after we realised it mattered — the shape of the answer itself.

The method has obvious limits and they are worth stating up front. This is one prompt, one engine, one metro, sampled once a week — not a representative survey of AI recommendations across industries. Phrasing is known to swing these results, and we deliberately held the phrasing constant rather than testing variants, which means we measured week-to-week volatility while holding wording fixed. We also have no visibility into the retrieval layer itself; we only see the sources the model chose to cite in its final answer.

What makes the log useful despite those limits is that it is longitudinal and unedited. Over fourteen weekly runs our own business was named five times and absent nine, including one unbroken five-week absence in mid-summer and three appearances in the most recent four weeks. An agency publishing only the favorable samples would produce a very different-looking dataset from the same experiment.

Does ChatGPT answer this question with prose or with a business panel?

Both, and which one you get decides whether your content matters at all. In a growing share of runs the model opens with a Google Business Profile panel of local firms before writing any prose, and every business in that panel carries a review count. A profile with no reviews is not ranked low in that block; it is not in it.

This is the single most consequential thing we have learned since the first version of this article, and it is invisible if you only record which names appear. In the 7 September 2026 run the model listed fourteen Utah agencies in the panel — with review counts of 86, 66, 57, 55, 50, 41, 23, 22, 17, 17, 4, 3, 1 and 1 — and then wrote a separate prose shortlist beneath it. Our business appeared in the prose shortlist and was absent from the panel, because our Google Business Profile has zero reviews.

Note where the review threshold actually sits. We initially assumed the panel required substantial review volume, on the strength of Trustpilot research reported by TechRadar Pro (2026) finding that businesses with 80 or more reviews were referenced in over 75% of relevant AI answers versus about 1% for those with none. The panels we have logged say something more useful: firms with a single review appear in them. Volume governs how often you are cited once you are eligible; eligibility itself appears to turn on having a rated profile at all. For a business sitting at zero, that reframes the job from a year-long review programme to getting the first three or four.

Which sources did the model actually cite?

The cited source mix rotated continuously across the fourteen weeks — independent blog roundups, then a narrow pair of agency directories, then agencies’ own service pages alongside academic and news coverage, then back to city-level directories. No single source type held for more than about three consecutive runs, which is the most practically important finding in the log.

In late July the model built its Utah shortlist from independent roundup posts — sites publishing “best GEO agencies” and “best SEO agencies in Provo” listicles — plus a business news article and one agency’s own site. Two weeks later the picture narrowed sharply: for that run, the model cited exactly two sources for its entire shortlist, both agency directories, one of them a “Top SEO Companies in Salt Lake City” ranking page and the other an AI-optimization directory listing. In the most recent run the mix changed again, this time drawing on individual agencies’ service pages, a Reddit thread discussing one agency’s reputation, an arXiv paper, and a report from The Verge (2026) on the AI SEO industry.

The practical read is that there is no single door to walk through. A business optimizing purely for directory presence would have been well positioned in one of these weeks and invisible in another. The source types that recurred most often across the whole period were third-party directories and roundups, followed by agencies’ own service pages when those pages were specific enough to answer the question directly.

Sample week	Dominant source type	Notable detail
Late July	Independent listicles and roundups	Several “best agencies” posts, plus one news article
Early August	Agency directories only	Exactly two directory sources for the whole shortlist
Mid August	Agency service pages, forum, research, news	Reddit reputation thread cited alongside an arXiv paper
Late August	Third-party city roundups	Review-ranked business panel led the answer for the first time
Early September	City directories plus two agency homepages	Panel of fourteen firms, then prose sourced from Expertise.com and Clutch

The two directories that recurred most often across the whole period were Expertise.com’s Provo list and Clutch’s Provo SEO rankings, both of which republish on a monthly cycle. If you are choosing one off-site placement to chase in this metro, the log points at those two before it points anywhere else.

Which agencies got named, and why those?

The named agencies changed considerably between runs, but the ones that recurred shared two traits: a service page that described AI visibility work concretely rather than as a slogan, and enough third-party footprint that the model encountered them outside their own domain. Longevity and review volume mattered less than we expected.

In the most recent sample the model’s first recommendation was an agency whose cited page laid out exactly what it does — traditional SEO combined with answer-engine work, citation tracking across named assistants, and specific coverage of Provo, Orem, Lehi and Utah County — along with a published starting price. The model quoted that page directly, including the price. The second recommendation was an established local firm whose about page explicitly described generative engine optimization. In both cases the model was not reading marketing atmosphere; it was extracting checkable facts from a page and repeating them.

That is a meaningfully different picture from the run two weeks earlier, where the model named ten agencies sourced entirely from directory rankings and quoted none of their sites. The lesson we take from the contrast is that a specific, factual service page is a genuine asset for this channel — it just is not sufficient on its own, because in some weeks the model never reaches any agency’s site.

One more detail is worth reporting because it cuts against the industry’s own marketing. When asked, the model volunteered a caveat unprompted: no legitimate agency can simply place a business into ChatGPT, because assistants decide what to mention based on available sources and signals. It then recommended asking any prospective agency for before-and-after examples showing AI citations rather than keyword rankings. That is better buying advice than most agency websites offer, including, until recently, parts of ours.

What does this mean if you’re trying to get recommended?

Build presence in the places the model reaches when it is not reading your site — directories, review platforms, and independent roundups — and make sure that when it does reach your site, the page states checkable facts rather than adjectives. Those are two different projects and most businesses fund only the second.

The first project is slow and mostly not a content exercise. Getting listed in the directories that publish city-level agency rankings, accumulating genuine reviews, and being named in independent roundups are all things that happen off your domain, on other people’s timelines. Our own log is the clearest evidence for why it matters: five consecutive weeks of absence, during a period when we were publishing steadily and improving our own rankings substantially. Content volume did not move this needle at all.

There is a real limit to how far off-site work alone gets you, though, and the panel finding above is it. Reviews are the one off-site signal that is simultaneously the cheapest to start, entirely within your control to ask for, and a hard gate on a whole answer format. Everything else in this section is a queue you join; reviews are a queue you can start today.

The second project is more tractable. The pages the model quoted had four things in common: they named specific engines rather than saying “AI”; they named specific geographies; they described a concrete process; and at least one published a real number. Vague pages did not get quoted even when the agency behind them was named from a directory. If you write nothing else this quarter, write one page about your service that a stranger could summarize accurately without calling you.

The third thing worth doing is simply measuring. Sampling costs almost nothing, and without it you are guessing about a channel that changes weekly. Ask the buyer question in your category, log the names and the sources, and repeat it on a schedule. You will learn more from four weeks of that log than from any general article about AI visibility, including this one.

What we’d actually do

We would treat the assistant shortlist as a lagging indicator of off-site reputation — but we are correcting the stronger version of that claim we made in August, because nine more weeks of data do not support it. We wrote then that the right move was to stop trying to move the shortlist with on-site work. Since then the model has named us in three of four runs and, in two of those, cited our own pages as its source — the Utah hub in one, the homepage in the most recent. On-site work is clearly reachable. It is just not sufficient, because in the weeks the model never leaves the directories, nothing on your domain is in the running.

The honest synthesis after fourteen weeks is a division of labour rather than a winner. Off-site reputation — directory listings, and especially review velocity, which gates the panel format entirely — decides whether you are a candidate. A specific, factual, checkable service page decides what the model says about you once you are. Fund both, in that order, and keep sampling: we will keep publishing the weekly log either way, including the weeks it says we are absent.

If you want to see what the assistants currently say about your business, book a call and we will run the prompts live on the call. For the on-site half of the work, see what an AI SEO agency in Utah actually does, or how we approach AI citation tracking.