How Directories Captured AI Local Recommendations

Here is the claim that sounds wrong until the data proves it: the four big AI engines agree on almost nothing, and that is exactly why directories run the table. ChatGPT, Claude, Gemini and Perplexity read largely separate slices of the web; we measured the gap at 83.7%, meaning more than four of every five sources an engine cited appeared in that engine and nowhere else. A category that fragmented should have no winners. It has one. Stop tracking individual domains, start tracking what kind of page keeps surfacing inside each engine's private slice, and the same shape appears in all four: a ranked list on a directory. The engines are reading different rooms of the same building, and a directory is sitting in every room.

That is the story this page tells, with first-party evidence rather than the usual advice. Most coverage of directories in AI local search is a how-to that ends at "claim your listings." This one starts from who actually gets cited and asks the harder question underneath: if aggregators have captured the answer layer across every engine, what happens to a small business that thought its own website was the thing being recommended?

How did directories quietly capture AI local recommendations across every engine?

Directories captured AI local recommendations by being the one page type that survives every engine's retrieval step, no matter how differently the engines are wired. An AI engine holds no private leaderboard of the best plumbers in your city. When the question lands, it fetches a small set of pages it can reach and trust, reads the business names inside them, and writes its answer from there. The page that wins opens with extractable, ranked, attributable names. A directory roundup is engineered for precisely that; a small business homepage, which opens with a tagline and a photo of a van, is not. So even though the four engines reach into four different source pools, the same artefact keeps coming back up in each pool, because it is the artefact best shaped for the job.

Two findings explain why the deck is stacked this way. Kevin Indig's early-2026 Growth Memo analysis of ChatGPT citations found that about 44 percent of AI citations come from the first 30 percent of a page: engines reward sources that answer the question early and plainly, which a directory titled "3 Best Roofing Contractors in Hamilton" does in its first screen. Indig's wider point, that engines cite genuine crawlable sources rather than inventing them, is the floor this story stands on; they are not hallucinating directories, they are choosing them. The peer-reviewed GEO paper by Pranjal Aggarwal, Vishvak Murahari, Ameet Deshpande and co-authors (arXiv:2311.09735, KDD 2024) supplies the second half: what lifts a source inside a generative answer is content-level signal, reviews, statistics, quotable structured detail, and targeted work on it can raise visibility by up to 40%. Page rank is not the lever. A directory is a machine for producing exactly those signals; most business websites produce almost none of them above the fold.

"Quietly" is the operative word. No engine announced it would defer to aggregators, and most owners never saw it happen, because the shift lives inside the retrieval step where nobody looks. You ask who the best HVAC contractor in your city is, you get five tidy names, and the directory that supplied four of them never appears in the sentence. The capture is invisible from the front of the answer and obvious from the back of it. The rest of this page is the back of it.

What does Formative Digital's scrape of 1,732 Ontario citations reveal about aggregator dominance?

It reveals that aggregators sit at or near the top of every engine's citation pile, even though the engines almost never cite the same individual page. We ran the study through DataForSEO's LLM endpoints against the live answer engines in May 2026, asking prompts of the form "who are the best {vertical} in {city}, Ontario" across nine Ontario cities, five trades, and four engines. The raw pull produced 176 successful queries, 1,732 citations, and 326 distinct domains; across the full comparison set of 583 domains over 44 city-and-vertical cells, only 95, or 16.3%, were cited by two or more engines, leaving 83.7% engine-private. We unpack that fragmentation in our cornerstone study on why four AI engines recommend four different shortlists for the same query. Here the point is narrower: inside each engine's private slice, a directory is usually the most-cited thing in it.

Look at what each engine reached for hardest. Claude's single most-cited domain in the entire scrape was the directory ThreeBestRated.ca, at 116 citations, followed by Custom-Contracting.ca, Yelp, the vertical aggregator Furnaceprices.ca, and Opencare. Perplexity's top sources were HomeStars, Opencare and BBB. Gemini routed nearly everything through vertexaisearch.cloud.google.com, its Vertex grounding wrapper, cited 384 times, and the named pages underneath that wrapper were disproportionately directories like Opencare, UrbanTasker and HomeStars. ChatGPT looks like the exception, leaning on google.com 130 times, until you remember that Google's local layer is itself an aggregation of profiles and reviews rather than the businesses own sites. Four engines, four different favourite directories, one shared habit.

The cross-city durability is the part that should change how an owner thinks. HomeStars and Opencare were each cited across all nine cities and by more than one engine, which is rare in a dataset where 83.7% of sources never cross an engine boundary. ThreeBestRated.ca appeared in eight of the nine cities. These are not lucky one-off hits. They are structural fixtures: whatever city you test, the same handful of aggregators keep showing up, which is the signature of capture rather than coincidence. By contrast, the individual business sites that did appear, a bergellaw.com here, a westwooddentalgroup.ca there, almost always showed up once, in one engine, in one city. The directory is the constant. The business website is the variable that mostly reads zero.

One worked example makes the abstraction concrete. For "best HVAC contractors in Mississauga," Claude cited UrbanTasker and Furnaceprices.ca; Gemini surfaced BBB, UrbanTasker, HomeStars and Furnaceprices.ca through its Vertex wrapper; Perplexity led with UrbanTasker and closed with HomeStars. Three engines, reading three different webs, and each one front-loaded a directory. The actual contractors, Martino HVAC, Maher Heating and Cooling, Applewood Air Conditioning, appeared as names inside those directory pages more reliably than they appeared as cited sources in their own right. That is the capture in a single query: the contractors are in the answer, but the directory is the source.

Want to see which directories are citing your competitors instead of you? Get a free AI visibility audit and we will pull your real per-engine citation profile from the same scrape behind this research.

Why does an engine cite a directory instead of your small business website?

An engine cites a directory instead of your website because the directory answers the exact question asked and your website answers a slightly different one. The query is comparative, "who is best," and a directory is a comparative document by construction: ranked, plural, review-weighted, with the names an engine needs sitting in extractable text near the top. Your site is a singular document. It argues that you are good, which is the right job for a human visitor and the wrong shape for a retrieval system trying to assemble a shortlist of five. The engine is not snubbing you. It is reaching for the page that is already formatted as the answer it has to produce.

Three structural advantages stack in the directory's favour, which is why this is a category outcome rather than a fixable quirk on one site. List structure: a directory opens with "Top 10" and delivers names immediately, satisfying Indig's first-30% rule that most SMB homepages fail. Review density: directories aggregate dozens or hundreds of reviews per business, the exact content-level signal the GEO paper found lifts visibility, and which a single site cannot credibly manufacture about itself. Perceived neutrality: an engine treats a third-party roundup as a more trustworthy judge of "best" than a business praising itself, the same instinct a careful human applies. The directory wins on format, on evidence, and on disinterest at once. Darren Shaw made the strategic version of this in Search Engine Land in May 2026: to perform in AI local search you have to shape what the broader web says about your business, because the engines scour third-party mentions rather than relying on your own site.

This is also why the four engines disagree on which directory while agreeing on the directory. Each engine trusts a different judge. Claude trusts the editorial shortlist, a human-curated ThreeBestRated.ca page where someone already chose the top three. Perplexity trusts breadth, your presence spread across HomeStars, Opencare and BBB at once. Gemini trusts whatever its Vertex pipeline can crawl and parse cleanly, which favours well-structured directory pages. ChatGPT trusts Google's structured local data, itself an aggregation. The judges differ, so the winning directory differs, but in every case the judge is a third party with a list, not your own homepage. We trace that per-engine split in detail in our look at how Perplexity blends directories and firm sites in its local answers, where the breadth pattern is clearest.

Is this an Ontario quirk, or the same pattern the public studies found?

It is not an Ontario quirk; three independent public studies found the same shape on different data, which is the strongest evidence that our local pattern is the general one. We triangulated our scrape against research we did not run, because a single agency's dataset proves little alone. When first-party Ontario data and three external studies on different geographies and methods all point the same way, the finding stops being ours and starts being the structure.

Read those four rows together and the logic locks. Yang's arXiv study supplies the mechanism at scale: engines read divergent source pools, yet within each pool citations concentrate on a few outlets. Our Ontario data names which outlets win the local long tail, the aggregators. BrightLocal shows the same outcome in US industries, dentistry being the cleanest case, an engine citing ten directories and zero practice sites for one vertical. Minneapolis Made supplies the tell that confirms the cause: the two engines that overlap most, Gemini and Perplexity, overlap because both are directory-heavy, pulling Thumbtack, BBB and Tripadvisor, so the shared core of AI local search is specifically the directory core. Simon Moser's Polygrowth study, reported in Entrepreneur, sizes the stakes: per BrightLocal, the share of people using AI to find a local business jumped from 6% to 45% in early 2026, which moves the answer layer the directories sit in to the centre of how owners get found. Profound adds why directories travel so well: each engine has a different default instinct, ChatGPT encyclopedic, Perplexity skewed to user-generated and directory content, Google AI Overviews more balanced, yet over 80% of all citations across every platform land on .com domains. Aggregators are .com publishers built to be cited; they fit every engine's instinct at once, which is why no single engine had to choose them for all of them to end up doing so.

What does directory capture mean for Brantford and Ontario SMBs, and what do you do about it?

For Brantford and Ontario SMBs it means your AI visibility increasingly runs through aggregators you do not own, and the response is to earn placement in them without betting everything on them. This is the uncomfortable middle. Ignoring directories cedes the answer layer entirely, because trade after trade the cited source is a directory, not a business site. Surrendering to them rents your visibility from a landlord who can raise the rent, reorder the list, or lose its own citation tomorrow. The work is both jobs at once: get onto the directories the engines read, and make your own entity legible enough that engines can also cite you directly. Own what you can, rent what you must, and never confuse the two.

This maps onto two of Formative Digital's 12 Vectors. Vector 5, Cite, is earning placement in the third-party sources each engine trusts, which in local search means the right directories for your trade. Vector 10, Localize, is making your local entity unambiguous to every retrieval system at once, so the same business resolves cleanly whether it is read through Google's Knowledge Graph, an editorial shortlist, or a review aggregator. We run both through the Formative Forces, our orchestrated multi-agent system, so a business is worked across all four source layers in parallel rather than one engine at a time. None of it is magic ranking dust; every step is verifiable against the same kind of scrape that produced these numbers, which is why we diagnose before we promise anything. A good place to begin is our guide to the local signals that decide who AI names in near-me answers.

The honest caveat belongs here, because this is money owners spend. Outcomes depend on your trade, your competition, and the digital presence you start from. Our own Brantford retail client Mattress Miracle grew from roughly 1,000 to more than 82,400 monthly organic visits (SEMrush, April 2026) through sustained structured-content work, and its owner Brad put it plainly: "In 40 years of advertising I've never seen anything like this. It's a completely new business." That is one industry and one starting point; yours will differ. The principle that travels is the structural one, not the number.

Which brings the argument to its edge. The danger is not that directories matter; it is mistaking rented ground for owned ground and stopping there.

"Listing on the directories is table stakes now, not a strategy. We watched four engines that agree on almost nothing all reach for the same aggregators, which tells you the directory is the rented front door to AI search. The mistake is to rent that door and call the job done. The work is to earn your place on the lists the engines read and to build your own entity so cleanly that the engines can cite you directly, so that the day an aggregator drops you, you are not invisible. Own the ground you can. Rent the rest with your eyes open."

- Matt Griffin, Founder, Formative Digital, Brantford, Ontario

Directories did not capture AI local search by being better businesses. They captured it by being better-shaped pages, ranked, reviewed, and quotable, in an answer layer that rewards that shape. SMBs who treat that as the whole game will spend the next few years renting visibility from landlords they cannot see. Those who treat it as one layer, and make their own ground legible underneath it, get named in four places while their competitors get named in one. For the regulated-trade version of this dynamic, where the directories and the advertising rules both tighten, read our study of how AI surfaces Ontario personal injury firms next.