Gemini names a specific franchise location 11.0% of the time, ChatGPT 1.2% — and AI answers get the details wrong about a third of the time

Asked a local question, Gemini names a specific franchise location 11.0% of the time, Perplexity 7.4% and ChatGPT 1.2%. That is close to a ninefold spread between assistants reading broadly the same underlying data. And the profile accuracy figure sits underneath all three: roughly one in three of the location details an AI assistant returns is wrong. Meanwhile the published channel data still puts Google Search as the first channel for 52% of local buyers. The correct posture is neither to ignore the surface nor to fund it as a media line, and most brands are currently choosing one of those two.

An assistant that names your location one time in ten and gets its details wrong one time in three is not a channel yet. It is a data-quality liability with a growing audience.

1. 01

   Accuracy is the actionable number; recommendation rate is not

   A brand cannot decide how often an assistant names it. It can decide whether the name, address, hours, phone and services the assistant repeats are correct. One of those two is a controllable input and the other is an outcome, and the model funds the controllable one.

2. 02

   The ninefold spread is a sourcing difference, not a preference difference

   The assistants are weighting different underlying records with different freshness. That makes per-assistant optimisation the wrong response and upstream record consistency the right one: fix the sources and all three improve, chase one assistant's behaviour and the work expires with the next model update.

3. 03

   Optimising for the assistant before the profile is complete is backwards

   The fields that decide 3-pack presence are largely the fields an assistant reads to answer a local question. A brand with incomplete categories and stale hours does not have an AI visibility problem yet, it has a listings problem that is now visible in a second place.

4. 04

   The budget discipline is the finding

   At recommendation rates between 1.2% and 11.0%, against a surface where a majority of local buyers still begin on Google Search, this is a hygiene programme rather than a media channel. The model sizes it accordingly and revisits the decision when the published rates move, not when the conversation does.

Location-mention rate: Gemini **11.0%**, Perplexity 7.4%, ChatGPT **1.2%** (~9× spread). AI-answer accuracy on location details: ChatGPT **68.3%**, Perplexity **68.0%** (≈1 in 3 wrong); Gemini **100%** in the same test. *(Corrected from a flat "⅓ of location data is wrong.")*

The model makes the canonical location record the single source and pushes it to every surface an assistant plausibly reads: the Google profile, the location page with structured data, the brand's own location directory, and the aggregators that resell the record onward. It then runs a scheduled prompt set per location against the three assistants and diffs the details each returns against the record. A discrepancy becomes a traced correction at whichever source produced it, rather than an appeal to the assistant, because the assistant is not the place the error lives. Reporting is on detail accuracy per location and per assistant, with recommendation rate tracked as context and explicitly not carried as a target.

1. 1

   Weeks 1–3

   Source map

   Every surface an assistant could be reading, catalogued and ranked by how widely it is resold

   Owner: OmniFlow

2. 2

   Weeks 3–7

   Record consistency pass

   Profile, location pages, structured data and aggregators aligned to the canonical record

   Owner: OmniFlow

3. 3

   Weeks 6–8

   Prompt set

   A fixed per-location prompt set built and run against all three assistants

   Owner: OmniFlow

4. 4

   Month 3 onward

   Diff and trace

   Returned details diffed, discrepancies traced to a source and corrected there

   Owner: OmniFlow

5. 5

   Month 3 onward

   Quarterly re-measure

   Accuracy re-measured per location and per assistant; recommendation rate tracked, not targeted

   Owner: OmniFlow

The model targets detail accuracy per location and reports it per assistant, because the three do not fail in the same places and an aggregate accuracy figure would hide which source is producing the errors. Recommendation rate appears in the report as context and carries no target, since it is set by model behaviour the brand does not control. Every published figure quoted here is a starting frame; the brand's own measured accuracy replaces it from the first quarterly run. Nothing above is presented as a forecast of how often a given brand will be named by any of the three assistants.

Modelled. Inputs: SOCi, Local Listings Benchmarks for Franchises 2026 (2,751 brands, ~350,000 US locations); BrightLocal, Where are your customers really searching? 2026 (1,227 US consumers). Modelled outputs are not a forecast or a guarantee of results. Cited: SOCi *Local Listings Benchmarks for Franchises 2026*.

Every figure below is published research, not a client result. They are the inputs to the arithmetic above, listed so it can be checked rather than taken on trust. The bracketed number points to the full citation at the end of this page.

These are the surfaces this engagement is run and measured from, shown with representative figures built around the benchmarks cited on this page. Every account we run reports into views like these, and you keep ownership of all of them.

Each figure on this page, the system it is read from, and the definition and window it is measured over.

Not applicable — this is a modelled engagement. Its weakest input is shelf life: recommendation rates are measured against a fixed prompt set at a point in time, on models that are updated continuously and without notice, so these three figures may not survive the year. The model treats them as a reason to fix location data rather than as a baseline to beat, and re-measures the brand's own prompt set quarterly instead of continuing to cite a published number after the first run.

A modelled capability, not a client account. Figures illustrate what the model produces and are labelled as modelled wherever they appear. See all our work

Full citations for everything cited on this page, with the sample and period each study covers, so you can go and read the original.