Auditing what AI assistants say about your franchise

A franchise candidate researching your brand increasingly starts by asking an assistant rather than a search engine. So does a customer looking for a local service. In both cases they receive a composed answer with a handful of citations, and in both cases that answer is being generated from sources you may not have looked at in years.

Most brands have never checked what that answer says. The ones that have checked usually did it once, informally, by typing their own name into one assistant — which is the least informative version of the exercise.

What follows is a process that produces something actionable rather than anecdotal.

Build the query set first

The audit is only as good as the questions. The failure mode is testing questions that presuppose the brand, because almost no real query does.

A franchise system has two distinct audiences and needs both covered:

The adversarial row is the one people skip and the one that most often produces a surprise. It is also the query a serious candidate absolutely will run before signing anything, so finding out what it returns is not optional.

Twenty to thirty queries covering these shapes is enough for a first audit. Fewer than that and you are sampling noise; many more and the exercise stops being repeatable, which defeats the purpose.

Run it properly

Assistants are non-deterministic. The same question asked twice produces different wording and sometimes different substance, and different assistants disagree outright. A single run tells you very little.

• Run each query at least three times, and treat only what recurs as signal
• Use more than one assistant. Coverage and citation behaviour differ enough that conclusions from one do not transfer
• Use a clean session with no personalisation or prior conversation, since your own history will skew results toward your brand
• Test from the right location where the query is local, because geography changes the answer substantially
• Record the date. These answers change, and an audit without dates cannot show movement

Capture three things per run: the substance of the answer, whether your brand appeared at all, and the full citation list. The third is the most useful and the most commonly discarded.

Classify what you find

Findings sort into four categories, and the category determines the remedy.

Absent

Your brand does not appear in unbranded category queries where it reasonably should. This is the most common finding and the most tractable. It generally means there is no source on the open web that answers that question while mentioning you — which is a content problem, not a reputation problem.

Stale

The answer describes your brand as it was. Old investment figures, a unit count from several years ago, a former positioning, locations that have closed. Almost always traceable to a third-party directory, an old press release or a franchise portal profile nobody has updated.

Wrong

Factually incorrect statements about the model, the territory structure, the support, or the requirements. Sometimes from a stale source, sometimes from a plausible-sounding inference, occasionally from confusion with a different brand.

Confused

Your brand conflated with another — a similarly named business, a former parent company, a franchise in a different category with the same word in the name. Multi-location brands get this more than most, and it is the category where structured data does the most work, as covered in schema markup for multi-location sites.

What you can actually fix

The uncomfortable part of this exercise is that not everything found is fixable, and the fixable things are not fixed by the action people reach for first.

1. Third-party sources you control or can request changes to. Franchise portal profiles, directory listings, your own old press releases, industry databases. This is the highest-yield work and it is unglamorous administrative effort
2. Gaps on your own site. If an assistant cannot state your investment range because no page states it plainly, the fix is a page that states it plainly. Most "absent" findings resolve here
3. Structured data and entity clarity, which addresses confusion findings more reliably than prose does
4. Content written to be quotable. A self-contained passage answering a question in two sentences is far more liftable than the same fact distributed across an article, which is the substance of getting cited, not just ranked
5. Nothing, for some of it. An unfavourable but accurate characterisation, or a wrong statement with no traceable source, may simply have to be addressed through the weight of better sources over time

That last item is worth being honest about internally before the audit runs, because the instinct on seeing a bad answer is to demand it be corrected today, and some of them cannot be.

The lag

Corrections do not appear immediately and the delay varies by mechanism. An answer composed from live retrieval can reflect an updated page within days. An answer drawn from a model's training data will not change until that model is retrained, which is outside your control and outside your timeline.

The practical implication is to fix the retrievable sources first — your own pages, the directories, the portal profiles — because those are the ones where effort produces visible change on a sensible horizon. Then re-run the audit on a schedule rather than watching for it.

Making it repeatable

A one-off audit produces a snapshot and an opinion. A repeated one produces a trend, which is what actually justifies continued effort.

Keep it deliberately light: the same query set, the same assistants, the same clean-session method, quarterly. Record presence, sentiment and citation sources per query. The useful outputs are whether you are cited more often than last quarter, whether the wrong answers you fixed have actually changed, and whether any new source has appeared in the citation lists — the last of which frequently reveals a directory or aggregator nobody knew was influential.

For a franchise system it is worth splitting the reporting by audience, because candidate-side and customer-side visibility are driven by different sources and improve at different rates. A brand can be well represented in local customer queries and entirely absent from franchise ownership queries, and the remedies have almost nothing in common.

Questions we get asked