Four assistants, four different local businesses

Executive summary

SOCi's 2026 Local Visibility Index measured how often multi-location brands get recommended by AI assistants for local queries, across more than 350,000 locations belonging to 2,751 brands with 50 or more sites each, scored on over 120 metrics across six platforms. The headline spread is wider than almost anyone in local search expects. Google's local 3-pack shows these brands 35.9% of the time. Gemini recommends 11%. Perplexity 7.4%. ChatGPT 1.2%.

A brand that is comfortably visible in the local pack is, by these numbers, roughly thirty times less likely to be named by ChatGPT. And the reason is not that its content is worse on one surface than the other. SOCi's own explanation points at profile data accuracy: 100% on Gemini, which is grounded directly in Google Maps, against 68% on ChatGPT and Perplexity. An assistant that cannot resolve which business you are, where it is, and whether it is open does not recommend it.

The second study makes the same point from the source side. Steady Demand analysed 14,472 citations from 1,487 local queries across 50 US metro areas and 10 service categories, published 19 August 2026. Gemini drew roughly 60% of its local citations from businesses' own websites, more than directories, review platforms and forums combined, with Reddit at 13.7%. ChatGPT leaned much more heavily on Reddit and directories. The two assistants cited the same domains 8% of the time. Not 8% more often than chance. 8%.

Then, in the same month that study published, Reddit blocked AI crawlers domain-wide.

The numbers, side by side

Sources: SOCi 2026 Local Visibility Index (recommendation rates, profile accuracy); Steady Demand, 19 August 2026 (source-diet figures for Gemini and ChatGPT).

Two caveats before anyone builds a strategy on this table. The SOCi index measures brands with 50-plus locations, so it describes chains and franchises rather than the single-location contractor or clinic. The direction almost certainly holds for smaller businesses; the magnitudes are not established for them. And the two studies used different prompt sets and different scopes, so the rows are compatible in direction and must not be arithmetically combined. That kind of cross-study combination is the error dissected in Every Study Agrees AI Citations Are Volatile.

Part 1 — Why Gemini is roughly ten times better at this

Gemini is grounded in Google Maps. That single architectural fact explains most of the gap, and it explains it in a way that is boringly actionable.

A local recommendation requires an assistant to resolve an entity: this business, at this address, in this category, with these hours, currently operating. Google Maps has spent two decades building exactly that resolution layer, with a verification loop, a review corpus and an ownership claim process attached. An assistant grounded in it inherits a 100% profile-accuracy baseline on the fields that matter, because the underlying record is canonical.

An assistant retrieving from the open web has to reconstruct the same record from whatever it can find: the business's own site, directory listings of varying age, review platforms, forum threads. SOCi puts the resulting accuracy at 68% on ChatGPT and Perplexity. Nearly a third of profile facts wrong or unresolvable is not a ranking disadvantage. It is a disqualification. A system that is not confident a business exists at a given address will simply not name it.

This is why the local AI problem is an entity problem before it is a content problem. More pages do not fix an unresolvable record.

Part 2 — The source diets are almost disjoint, and that is the strategy

An 8% domain overlap between ChatGPT and Gemini on local queries means these are not two views of one system. They are two systems.

Gemini's 60% business-website share says something unusually encouraging. For the assistant with the highest local recommendation rate, the most-cited source type is the business's own site. Not a directory. Not a review aggregator. Your own pages, presumably corroborated against the Maps record. For once the advice is simple: the asset you fully control is the asset that surface uses most.

ChatGPT's diet is the mirror image. Reddit and directories, sources the business does not own and cannot directly edit. Winning there historically meant being talked about in third-party spaces, which is a public-relations and community problem wearing an SEO costume. That is the same conclusion five independent studies reached about AI visibility generally, traced in SEO Is Becoming Brand Engineering, Not Ranking Engineering.

The practical consequence is that a single "AI visibility" programme for a local business is incoherent. The work that moves Gemini, meaning an accurate Maps record and structured current information on your own site, barely touches ChatGPT. The work that moved ChatGPT, meaning presence and reputation in third-party discussion, barely touches Gemini. Budget them separately or accept that you are optimising for one and hoping about the rest.

Part 3 — The Reddit hole in the middle of ChatGPT's local answers

Here is where the two 2026 stories collide.

Steady Demand published on 19 August 2026, measuring a period in which ChatGPT relied heavily on Reddit for local citations. Reddit blocked AI crawlers domain-wide in early August. Its share of ChatGPT Search citations fell from 3.83% to 0.52% between 7 and 17 August, an 86.4% relative drop, as ChatGPT's retrieval shifted toward domain-scoped site: queries against named, official sources. That episode is documented in Reddit Blocked the Crawlers.

A category of query that depended disproportionately on one source just lost that source. Nobody has published what replaced it for local specifically, and the honest position is that this is currently unmeasured. Two possibilities are worth naming, because they point in opposite strategic directions. If ChatGPT's local answers shifted toward directories, then directory hygiene, the unglamorous NAP consistency work, just got more valuable on that surface. If they shifted toward official business websites and institutional sources, following the site: fan-out pattern, then ChatGPT's local diet moved toward Gemini's and the 8% overlap figure is already out of date.

Either way, any local AI visibility baseline measured before August 2026 should be treated as historical. Anyone still quoting pre-August ChatGPT local benchmarks is describing a system that no longer exists.

Part 4 — What this means for a single-location business

The SOCi index measures 50-plus-location brands, and a Toronto renovation contractor or a single clinic is not that. The mechanics still transfer, because they are mechanics rather than scale effects.

Entity resolution is the floor. A business with one verified, complete, current Google Business Profile, consistent name-address-phone data everywhere it appears, and structured data on its own site that agrees with both is resolvable. One that has two half-claimed listings, an old suite number on three directories and no LocalBusiness markup is not, and no amount of content fixes it. On the Osoba Renos build, consolidating a scatter of listings into one entity and shipping LocalBusiness and Review schema preceded page-one positions for non-branded queries like "basement renovation Mississauga". The entity work came first because nothing downstream works without it.

Then the asymmetry works in a small business's favour on exactly one surface. Gemini takes 60% of its local citations from business websites, and a single-location business can make its own site the most accurate, most current, most specific source about itself far more easily than a 400-location chain can. That is a real edge, on the assistant with the second-highest recommendation rate, available to anyone willing to do unglamorous data work.

Counter-argument, taken seriously

ChatGPT's 1.2% may be measuring the wrong thing. The strongest objection is that ChatGPT is not primarily a local discovery product and was never trying to be. People asking it for a contractor are a minority use case against a product built for reasoning, drafting and general questions. A 1.2% recommendation rate for a surface nobody uses that way is a small number attached to a small denominator, and optimising hard for it could be chasing volume that does not exist.

That objection has real force, and it changes prioritisation rather than the finding. Two things push back on it. Local intent inside general assistants is growing rather than shrinking, and businesses that establish resolvable entities before that traffic matures will not be easy to displace later. And the underlying fix, an accurate verified consistently-described entity, is the same work that moves Gemini, the local pack and Perplexity at the same time. It is not a ChatGPT-specific investment. The correct reading is to do the entity work for the surfaces that already send business, and treat ChatGPT presence as a free option the same work buys.

What everyone is missing

This is a data-quality problem being sold as a content problem. A 32-point profile-accuracy gap between Gemini and ChatGPT is the clearest explanatory variable in the SOCi data, and it is about records, not prose. Most local GEO advice on sale in 2026 is content advice. The binding constraint is upstream of content.

Nobody is measuring single-location businesses at scale. Every substantial dataset here covers multi-location brands with 50-plus sites, because those are the accounts that buy visibility software. The overwhelming majority of local businesses sit outside every published figure in this field, and the assumption that the magnitudes transfer downward is an assumption, not a finding.

A 35.9% local-pack rate against an 11% best-case AI rate means classic local SEO is still where the volume is. That is not a fashionable conclusion in a GEO piece, and it is what the numbers say. The local pack remains the highest-yield local surface by better than three to one over the best assistant. The case for doing the AI work is that it is largely the same work, not that it has overtaken anything.

Cross-engine overlap of 8% means "we rank in AI" is not a sentence. Being recommended by one assistant predicts almost nothing about the others. Any local visibility report presenting a single blended AI number is hiding four different results inside one average, for the same reason set out in the volatility piece above.

Future predictions

• ChatGPT's local recommendation rate rises as its source diet shifts off Reddit. Domain-scoped retrieval against official sources is structurally better for local resolution than forum threads were, even though the transition cost real citation volume.
• Profile accuracy becomes the headline local AI metric. It explains more variance than anything else currently measured and it is directly actionable, which is the combination that makes a metric win.
• Someone publishes a single-location dataset within a year, and the recommendation rates come in lower than the chain figures, because chains have dedicated listings management and independents do not.
• Google's Maps grounding advantage widens before it narrows. No competitor has an equivalent verified local record, and building one is a decade-scale undertaking rather than a model improvement.

Practical takeaways

1. Fix entity resolution before writing anything. One verified Google Business Profile, one canonical name-address-phone string, byte-identical everywhere it appears. This is the constraint, and it sits upstream of every other tactic. See Local SEO in Toronto.
2. Put the facts an assistant needs on your own site, in structured form. Gemini takes about 60% of local citations from business websites. LocalBusiness schema with address, hours, service area and services, agreeing exactly with your Maps record.
3. Treat each assistant as a separate surface with its own baseline. At 8% overlap, one blended AI visibility number is not a measurement. See LLM Visibility Monitoring.
4. Re-baseline any local AI measurement taken before August 2026. Reddit's crawler block changed ChatGPT's local source mix mid-period. Pre-August benchmarks describe a system that no longer exists.
5. Keep the majority of local budget on the local pack. 35.9% against 11% is not close. The AI work is justified because it is the same entity work, not because it has overtaken classical local search.
6. Audit directory listings for accuracy, not volume. 68% profile accuracy on the open-web assistants is a data-hygiene number. Fewer, correct, current listings beat more listings.
7. If you serve a defined metro, say so explicitly and consistently. Ambiguous service areas are one of the commonest reasons an assistant declines to name a business for a place-qualified query. See GEO agency in Toronto.