Why AI Engines Recommend Different Local Businesses

The consensus gap, by source layer (Formative Digital scrape, 1,732 citations, May 2026)

Engine	Source it leaned on hardest	Times that domain was cited	Share of its citations unique to it
ChatGPT (OpenAI)	google.com	130	83.7% of all cited domains appeared in only one engine
Claude (Anthropic)	threebestrated.ca	116
Gemini (Google)	vertexaisearch.cloud.google.com	384
Perplexity	homestars.com	17

Read that table as four separate maps of the same city drawn by four cartographers who never spoke to each other. ChatGPT is essentially reading Google. Claude is reading one Canadian directory editor's shortlist. Gemini routes nearly everything through its own Vertex grounding pipe, which it cited 384 times, more than any actual publisher in the dataset. Perplexity scatters across review sites. The primary keyword question, why AI engines recommend different businesses, has a structural answer rather than a mysterious one: they are grounding in source layers that barely touch. Below is the full evidence, three side-by-side examples you can check yourself, and what the gap means whether you are choosing a business or trying to be the one chosen.

If four engines answer the same question, why do you get four different shortlists?

You get four shortlists because each engine assembles its answer from a different pool of sources, fetched fresh at the moment you ask. An AI engine holds no internal leaderboard of the best dentists or roofers in your city. When the question arrives, it runs a retrieval step, grabs a small set of pages it can reach and trust, reads the business names inside them, and writes its list from there. Swap the source pool and the list changes. The four major engines have each made durable, distinct choices about which pools to draw from, and those choices are consistent enough that we can fingerprint them, which is exactly what the opening table does.

This is the move most coverage of the topic skips. Plenty of articles list the generic reasons engines differ, things like different training data, different temperature settings, different indexes, and then stop at the abstraction. A smaller set runs one cute anecdote, such as Yext asking four engines where to get the best martini and fries in New York and getting four different answers with zero overlap. The anecdote is directionally right and we will return to it, but a single query proves very little. What the field has lacked is a measurement of the gap at scale, in the messy local long tail where it bites hardest. So we built one.

The mechanism underneath has a name in the research. Pranjal Aggarwal, Vishvak Murahari, Ameet Deshpande and their co-authors set it out in their paper "GEO: Generative Engine Optimization" (arXiv:2311.09735, presented at KDD 2024). Their core finding is that what surfaces a source inside a generative answer is content-level and citation-level signal, reviews, statistics, quotations, crawlable structured detail, and that targeted work on those signals can lift a source's visibility by up to 40 percent. Page rank is not the lever they found. Citable, attributable, well-structured content is. That single fact is why the same business can win one engine and vanish from another: each engine is grounding against a different body of citable material, and the material disagrees.

Kevin Indig's February 2026 Growth Memo, built on 1.2 million AI answers and 18,012 verified ChatGPT citations, sharpens this with a positional rule that runs through everything below: 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. A business that buries its city and its services under a paragraph of brand storytelling hands the engine nothing extractable near the top, so the engine reaches instead for a directory that opens with a clean ranked list. Indig's work also established, on real data, that these engines cite genuine crawlable sources rather than inventing them. Our scrape extends that principle into local service queries and adds the comparison Indig did not run: not just what one engine cites, but how little four engines agree with each other.

How much do the engines actually overlap, and where did 16.3% come from?

The engines overlap on 16.3% of their sources, which is the share that appeared in two or more of the four engines, and that number came straight out of our citation database rather than from any estimate. We ran the scrape through DataForSEO's LLM endpoints against the live answer engines in May 2026. The job produced 176 successful queries, 1,732 individual citations, and 326 distinct domains in the raw pull, every one a real local-intent prompt of the form "who are the best {vertical} in {city}, Ontario." For the cross-engine comparison we widened to the full comparison set of 583 distinct domains spread across 44 city-and-vertical cells, then asked the plain question: how many of those domains did two or more engines cite? The answer was 95. Ninety-five out of 583 is 16.3%, which means the other 83.7% of every source an engine cited lived inside that one engine's answers and nowhere else.

We measured the agreement with Jaccard similarity, the size of the intersection of two engines' cited-domain sets divided by the size of their union. A Jaccard of 1.0 would mean two engines cite an identical source set; 0.0 means they share nothing at all. Every pairwise figure in our data sits far closer to zero than to one. The 16.3% headline uses a deliberately generous bar, cited by two or more engines. Raise the bar to cited by three or more, and the shared core collapses to a handful of large aggregators, mainly homestars.com and opencare.com. Four engines almost never line up on the same source three deep.

This is the resolution the national studies cannot reach, and it is why our number lands where it does. BrightEdge, in its April 2026 study of five engines across nine industries, found pairwise top-100 cited-source overlap ranging from 16 to 59 percent, a 43-point spread, while brand-recommendation overlap held tighter at 36 to 55 percent. Our local figure sits at the very bottom of BrightEdge's range, which is exactly what theory predicts. National brand queries have a dense, heavily linked core of sources every engine already knows, so the overlap drifts upward. Local service queries have no such core. The authoritative source pool for "best HVAC contractor in Mississauga" is thin, so each engine's individual grounding choice dominates the result, and the agreement falls away. Tinuiti's Q1 2026 AI Citation Trends Report, covered by Search Engine Land, reached the same conclusion from a different angle: there is no universal top source for brands, and what gets cited on Perplexity has almost nothing to do with what works on Gemini.

The cleanest way to feel the gap is to stop looking at aggregate percentages and watch it happen inside three single prompts. We pulled three real queries from the scrape and recorded the businesses each engine named. Start with Toronto dentists. ChatGPT, reading google.com, returned The Richmond Dental Centre, Chaplin Dental, Clinton Dental, The Dentistry Place, and Portrait Dental. Claude, reading curated lists, returned Dr. Judy Sturm and Associates from yorkvillesmiles.com, Metropolitan Dental, Artin Dental Clinic, and an Opencare shortlist. Gemini, wrapping its sources through Vertex AI Search, surfaced opencare.com, streetsoftoronto.com, donddental.ca, therichmonddentalcentre.com, and a Reddit thread. Perplexity returned an Opencare top-rated list, Bite Dental, hellodent, Dr. Judy Sturm again, and a 123Dentist page. Exactly one business, The Richmond Dental Centre, appears in two of the four lists. Everything else is unique to a single engine. A Toronto dentist proudly screenshotting their ChatGPT placement is invisible inside Claude and Perplexity for the very same question.

Now Hamilton roofers, where the lists are even more disjoint. ChatGPT named Silva's Roofing and Siding, Complete Exteriors, The Roofing Master, Boucher's Roofing, and Edge's Roofing Co, all via google.com. Claude pulled a threebestrated.ca shortlist, The Roofing Crew, a Melanie Jade Design "7 best" roundup, and a HomeStars Hamilton page. Gemini surfaced gaf.ca, dangeloandsons.com, guildquality.com, roofing.ca, and professionalroofers.com, leaning on manufacturer and directory sites rather than local firms at all. Perplexity returned a HomeStars Hamilton page, Platinum Roofing, Capela's Roofing, DeLuca Roofing, and a GAF contractor list. No single roofing company is named by all four engines. Claude and Perplexity touch only at the HomeStars listing page, not at a specific contractor. Four engines, four realities. The pattern holds for Mississauga HVAC too, where UrbanTasker bridges Claude and Perplexity and furnaceprices.ca bridges Claude and Gemini, yet not one contractor, not Martino HVAC, not Maher Heating and Cooling, not Applewood Air Conditioning, is named by every engine. The more local the query, the less the engines agree. That is the rule, and it repeats across every city and vertical we tested. We break the dental vertical down on its own in our companion study on how AI engines choose which Ontario dental clinics to surface, where the directory mix shifts again.

Why does ChatGPT echo Google while Claude trusts a directory editor?

ChatGPT echoes Google and Claude trusts a directory editor because the two companies wired their retrieval layers to fundamentally different sources, and those wiring choices show up cleanly in the citation data. Each engine has a stable, recognisable set of places it reaches for first, what we call its grounding fingerprint. Knowing the fingerprint is the difference between optimising in the dark and optimising on purpose, so it is worth walking through all four.

ChatGPT reaches for google.com first, 130 times in our scrape, and then for individual business websites such as bergellaw.com, westwooddentalgroup.ca, and wmgreenroofing.ca. In practice that means ChatGPT is reading Google's structured local data, the Maps listings, the reviews, the Knowledge Graph, and pairing it with whatever it can pull from a business's own site. Owned property matters most here. Whatever Google knows about your business flows fairly directly into ChatGPT's answer, which is why its picks often read like a local pack with full sentences attached. It is also why a strong, accurate Google Business Profile is the single most useful asset you can fix for this one engine.

Claude is the most editorial of the four. It leaned hardest on threebestrated.ca, cited 116 times, followed by custom-contracting.ca, yelp.com, furnaceprices.ca, goodcaring.ca, and opencare.com. These are human-curated directories and "best of" lists, places where an editor has already done the work of choosing a shortlist. Being merely present is not enough for Claude; being editorially selected onto the list is what gets you named. A business that has earned its spot on a curated directory has a real advantage inside Anthropic's engine and almost no advantage inside ChatGPT's google.com layer, which is the whole point.

Gemini is the strangest line in the data and deserves a careful look. It cited vertexaisearch.cloud.google.com 384 times and almost nothing else by name. Vertex AI Search is Google's grounding service, so what you are seeing is the redirect wrapper rather than the publishers underneath it. For a business owner, the practical consequences are twofold. First, Gemini's true sourcing is the hardest of the four to inspect, because the actual pages are masked behind the Vertex pipe. Second, the lever you do have is crawlability and structure: if Google's grounding pipeline can fetch, read, and parse your content cleanly, you are in the running, and clean Schema.org markup is how you make that fetch reliable. You cannot see the layer, but you can feed it.

Perplexity spreads broadly across review aggregators and rating bodies, homestars.com (17), opencare.com, and bbb.org, alongside recurring firm sites like preszlerlaw.com and mcleishorlando.com. Breadth of presence is the currency here. Perplexity rewards businesses that show up across many review surfaces rather than dominating any single one, which makes it the engine where a wide, consistent footprint across the review web pays off. Notice that opencare.com and homestars.com are two of the very few domains that surface across more than one engine; the aggregators with genuine cross-engine reach are precisely the ones that crack the 16.3% shared core.

Set the four side by side and the headline reasserts itself. ChatGPT optimises for Google's structured data, Claude for editorial selection, Gemini for crawlable structure behind an opaque wrapper, Perplexity for breadth across reviews. These are not four flavours of one system. They are four systems wearing similar chat windows. The same divergence is visible at the level of two engines in our closer look at how ChatGPT and Perplexity behave as different answer engines, and the strategic gap it opens up, where a Google ranking does not equal an AI citation, is the subject of our piece on GEO versus SEO in 2026.

One reasonable worry at this point is whether all this divergence simply means AI recommendations are random. The honest answer is no. The gap is divergence, not randomness, and the difference decides whether optimisation is worth attempting. Each engine is internally consistent about which layer it trusts: ChatGPT reliably reaches for google.com city after city, Claude reliably reaches for curated directories. That is structure, and structure can be engineered against. There is a separate, genuine source of run-to-run noise on top of it. SparkToro, in research led by Rand Fishkin, had 600 volunteers run 12 prompts across ChatGPT, Claude, and Google AI a combined 2,961 times and found that an engine asked the same question 100 times has a less than one in one hundred chance of returning the identical brand list in any two responses. So two true things stack: the engines differ from one another systematically, and each engine wobbles a little from run to run. The structural divergence is the large, addressable effect; the wobble is the small one. Neither is random in the sense of being beyond your influence.

If you are choosing a business from an AI answer, which engine deserves your trust?

If you are choosing a business, no single engine deserves your whole trust, and the smartest move is to treat any one AI answer as one opinion rather than a verdict. Because 83.7% of cited sources are unique to one engine, the list you get is shaped as much by which engine you happened to open as by which businesses are genuinely good. Ask ChatGPT and you are largely reading Google's local data. Ask Claude and you are reading a directory editor's shortlist. Ask Perplexity and you are reading a synthesis of review sites. Each is a real signal, and each is partial. Cross-checking two or three engines is the consumer version of getting a second quote.

Accuracy and coverage are also two different things, and it pays to keep them separate when you are deciding who to call. On raw factual accuracy of business details, Gemini currently leads because it grounds in Google Maps. Search Engine Land's 2026 AI Local Visibility Report found business-profile information was roughly 100 percent accurate on Gemini, against about 68 percent on ChatGPT and Perplexity. That sounds like a clean win for Gemini until you look at coverage. The same report found that across local-business queries, ChatGPT recommended only 1.2 percent of locations, Gemini 11 percent, and Perplexity 7.4 percent. Every engine surfaces a far narrower slice of the local market than Google's own Maps does, which means the genuinely good business two streets over may simply not appear in any AI answer yet, not because it is worse, but because it has not yet earned a place in the source layers these engines read.

There is also a behavioural risk worth flagging, because it changes how much weight you should put on any AI recommendation. People over-trust a confident synthesised answer precisely because it arrives clean, sourced-looking, and free of the ten blue links you would normally scan. Five dentists in tidy prose can feel more authoritative than a Maps list of forty, even though the Maps list is more complete and the engine's five were chosen by whichever sources it could ground. For a low-stakes decision that is fine. For anything consequential, a roofer who will be on your home for a week, a clinic you trust with your health, the AI answer is a starting shortlist, not a verdict. Open a second engine, then check the businesses against their actual reviews and credentials. The consensus gap is, in effect, a built-in warning label: when four engines disagree this sharply, no one of them is the final word.

If you are the business, how do you get named by engines that read different webs?

If you are the business, you get named by earning a place in each engine's preferred source layer at the same time, rather than betting everything on the one you happen to understand. Because the four layers are known and stable, this is concrete work, not a guessing game. The GEO paper found that targeted content optimisation can raise a source's visibility inside generative answers by up to 40 percent, and that the effective tactics are domain-specific, which is the academic way of saying the work is real and the tactics differ per engine. Here is how the four fingerprints translate into action.

For the ChatGPT layer, the lever is google.com. That means a complete and accurate Google Business Profile, consistent name, address, and phone data everywhere it appears, and a steady flow of genuine Google reviews. This is the one layer where your owned Google presence converts most directly into AI visibility, so it is usually the first thing we fix. For the Claude layer, the lever is editorial selection. Earn placement on human-maintained shortlists such as threebestrated.ca, and on category directories like opencare.com or furnaceprices.ca, and understand that being chosen onto the list, not merely listed somewhere, is what counts. For the Perplexity layer, the lever is breadth. Build a consistent, accurate presence across homestars.com, opencare.com, and bbb.org, because Perplexity rewards businesses visible across many review surfaces rather than dominant on one. For the Gemini layer, the lever is structure. Publish crawlable, well-organised content with clear Schema.org markup so Google's Vertex grounding pipeline can read and cite it, since structured data is the only real handle you have on a layer you cannot inspect directly.

This maps onto two of Formative Digital's 12 Vectors in particular. Vector 5, Cite, is the discipline of earning placement in the third-party sources each engine trusts. Vector 10, Localize, is the discipline of making your local entity unambiguous to every retrieval system at once, so the same business resolves cleanly whether it is being read through Google's Knowledge Graph, a directory shortlist, or a review aggregator. We run both through the Formative Forces, our orchestrated multi-agent system, so a single business is worked across all four source layers in parallel instead of one engine at a time. The reason this is not a magic-ranking-dust pitch is that every step above is verifiable against the same kind of scrape that produced this article's numbers; if you want a starting read on your own position, the soft first step is a diagnosis of where each engine currently surfaces you, which is Vector 1.

Measurement is where most businesses go wrong, so it gets its own paragraph. If 83.7% of cited sources are unique to one engine, then checking "do I appear in ChatGPT" tells you almost nothing about the other three, and because each engine wobbles run to run, a single check on a single day tells you little even about that one engine. Honest measurement is cross-engine and repeated. Track your presence in all four engines as four separate scores, not one blended "AI rank," and sample each engine several times to average out the run-to-run noise SparkToro documented. Watch the source layers as well as the answers, because the layer is the leading indicator and the citation is the lagging one. If you have just earned a spot on threebestrated.ca, expect movement in Claude before you see it anywhere else.

Why does any of this matter commercially? Because the audience asking these engines is no longer a rounding error, and it is compounding. Statistics Canada reported that 12.2 percent of Canadian businesses used AI to produce goods or deliver services in the twelve months to the second quarter of 2025, double the 6.1 percent of a year earlier. A business absent from three of four engines is absent from a growing share of how people now pick a dentist in Brantford, a roofer in Hamilton, or an HVAC contractor in Mississauga. One honest caveat belongs here: outcomes depend on your industry, your competition, and your existing digital presence, and this work is not a switch that flips on identically for everyone. 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 as their owner Brad put it, "In 40 years of advertising I've never seen anything like this. It's a completely new business." That reflects one industry and one starting point. Yours will differ, which is why we diagnose before we promise anything.

Step back from the tactics and the shape of the whole thing is simple, which is also why it is so easy to get wrong. People keep asking how to rank on AI, as if there were one ranking to win.

"There is no single AI ranking. We measured four engines reading four different slices of the web for the same Ontario query, and 83.7 percent of what they cited appeared in only one of them. So the work is not chasing one algorithm. It is making your business legible to four separate retrieval systems at the same time, then measuring each one on its own terms. That is engineering, not magic ranking dust, and it is the whole reason a well-built local business can show up in four places while its competitors show up in one."

Matt Griffin, Founder, Formative Digital, Brantford, Ontario

The consensus gap is not a problem to be solved so much as a structure to be worked with. Four channels, four rankings, one chat box. Measure it that way and optimise it that way, and the gap that buries businesses who treat AI search as a single ranking becomes the opening for one that does not. This page is the cornerstone for that argument; the regulated-vertical version, where the same divergence plays out under professional advertising rules, is worth reading next in our study of which Ontario injury-law firms the four engines actually name.