What agencies get wrong reporting AI visibility to franchise and multi-location clients

Picture an illustrative franchise client with fourteen locations who gets a monthly report showing “AI visibility: 22%, up three points.” It looks like steady progress. What it does not show is that two locations are named constantly, six are named occasionally, and six have never been named once. Averaged together, that looks like a brand doing okay. Location by location, it is a brand where more than a third of the business is invisible to AI answers, and the client has no way to know that from the number they were handed.

This is the reporting mistake specific to multi-location and franchise accounts, and it is different from the mistakes covered in reporting AI visibility to clients generally. A single-location client has one entity to explain. A franchise client has a portfolio, and a portfolio needs a report built to show variance, not to erase it.

Why one number breaks the moment a client has more than one location

AI answers are already far more selective than the ranked lists franchise clients are used to reading. SOCi’s 2026 Local Visibility Index, built from roughly 350,000 locations across 2,751 multi-location brands, found ChatGPT named only 1.2% of individual locations analyzed, Perplexity named 7.4%, and Gemini named 11%, against 35.9% visibility for the same brands in Google’s local 3-pack - up to 30 times more selective than Google’s local results, by the study’s own comparison.

A shortlist that tight does not distribute evenly across a brand’s footprint. Some locations clear the bar consistently, most do not, and a blended percentage puts both groups inside the same digit. Going back to the illustrative “22%” report above: the client reading it cannot tell whether that is fourteen locations each doing okay, or two locations carrying the entire number. Those are different businesses with different problems, and they need different reports.

What the average is actually hiding

The variance is not random noise, it traces back to per-location data quality, and that is measurable. SOCi’s franchise benchmarking found that even at brands with near-universal Google Business Profile claim rates, only about 80% of locations listed on Google also have a Yelp page, and just 53.4% have a matching Facebook page, with profile completeness running 86.6% on Google, 81.3% on Yelp and 87.4% on Facebook. Those gaps are never spread evenly. A handful of locations opened under a different manager, inherited from an acquisition, or simply never finished setup, and they sit well below the brand’s own average while a few flagship locations sit well above it.

The same benchmarking traced why engines differ so much in how often they name a location: business contact information pulled by ChatGPT was only 68.3% accurate and Perplexity’s 68.0%, against 100% for Gemini, which is grounded directly in Google Maps data rather than a mixed crawl of listings. An engine that cannot confirm a location’s hours or address with confidence has an easy, safe answer available: name a different location, or name a competitor instead. That is a data-quality story, not an AI-visibility story, and a blended score cannot tell a client which one they are looking at.

The mistake that is not the average: reacting to noise

The second failure mode shows up after an agency starts reporting by location instead of blended, and it is the opposite problem: treating every location-level swing as a finding. AI answers are noisy at the level of an individual question, and the noise gets worse the narrower the slice you look at. An analysis of 851 tracking runs across 81 brands, with no changes shipped between consecutive checks, found overall visibility scores moved by up to 14.5 points at the 95th percentile, but a single category-level cut swung by up to 27.3 points - nearly twice as noisy as the overall number, because a narrower slice has fewer answers behind it and a single flipped answer moves it further. A single location’s result is an even narrower slice than a category, which means it is the least stable thing in the entire report and the easiest one to over-read.

The same analysis found individual prompt-engine results were actually stable, flipping only 6.4% of the time between runs - the instability is mostly arithmetic on small numbers, not the engines being genuinely fickle. Its fix is worth adopting directly: a first sighting of a mention survived to the next run only 54.3% of the time, but a mention confirmed on two consecutive runs survived 96.1% of the time. Don’t tell a client a location “got named” off one good check, and don’t tell them a location “lost” its answer off one bad one.

What to report instead of a single score

Two changes fix most of this, and neither requires new tooling beyond what a location-level view already gives you.

Report the distribution, not the average. Show the median location’s presence rate alongside the range: best location, worst location, and the count of locations below whatever floor the client considers acceptable (zero mentions across the tracked question set is the obvious one). To go back to the illustrative account from the opening: “median 24%, six of fourteen locations at zero” tells that client something true that a single blended average does not. Measuring share of voice covers the mechanics of the underlying number; for a multi-location client, compute it per location first and roll up second, never the reverse.

Separate the entity problem from the content problem. A location stuck at zero because its Yelp listing is missing and its hours are wrong needs a different queued action than one stuck at zero because its service pages are thin. Getting a local business recommended by AI covers the checklist to run per location, and it is worth running it location by location rather than assuming the brand-wide fix already covered every branch. A single push at headquarters rarely reaches the location that needed it most.

Explain the variance instead of smoothing it over

The instinct to blend scores usually comes from wanting to avoid a hard conversation: “why is location nine still at zero after three months.” The honest answer is almost always more defensible than the blended number that hides the question. If nine’s Facebook page was never claimed, or its GBP category is wrong, that is a specific, fixable cause you can point to and a specific date you fixed it. Different engines will keep disagreeing about which locations they trust regardless of the work you do, which is exactly why tracking each engine separately matters more for a multi-location account than a single-location one - a location that Gemini trusts and ChatGPT does not is telling you something real about that location’s Google Business Profile versus its wider listing footprint, and averaging the two engines together erases the clue.

It also helps to set expectations before the first report goes out, not after a client questions the second one. Tell them up front that location-level results will vary more than the brand-wide number suggests, that a single check on a single location proves close to nothing, and that the report will show the spread on purpose rather than a single reassuring digit.

Build the roll-up so it survives scrutiny

None of this is more work than the blended report, it’s the same underlying data cut a different way: by location before it’s compressed into a brand number, with the two-run confirmation rule applied before anything counts as a win or a loss. If you run AI visibility across several multi-location or franchise accounts, keeping the location-level breakdown as the default view rather than something you compute on request is what makes the harder, more honest report the one you actually send every month. See how this works for agencies managing several accounts, or run a free Visibility Check against a client’s flagship and one struggling location side by side to see the gap this piece describes, before you build the first report around it.