Ask Gemini which payroll platform suits a 40-person company. Then stop reading the answer and look at what built it.
The sources will not match your Google results. Some will be review platforms, some a forum thread, some a YouTube explainer nobody on your team has watched.
Heather Campbell made the operational version of this point in Search Engine Journal this week. The sources shaping AI answers in your industry form their own set, and mapping it is separate work from checking where you rank.
That sounds obvious. Almost nobody does it.
Your category picks the source mix, not you
Campbell's data shows the mix changing sharply by industry. Consumer retail answers lean on structured product information and brand-owned pages. IT services answers lean on review sites like G2 and on community platforms.
Two marketing directors with the same budget and the same competence therefore face two different jobs. One needs product data that is clean and machine-readable. The other needs a presence in conversations it does not host.
Most category strategies get copied sideways from whoever published the loudest case study. Take a retail playbook into a B2B services company and you will produce immaculate product schema for an answer that is being assembled out of G2 profiles and Reddit threads.
Community and user-generated content keeps appearing across consumer products, IT solutions and communication services. Reddit and YouTube are not a channel you bolt on at the end. In several categories they are the raw material the answer is made of.
They are also the one place where standard marketing instincts actively hurt. Disguised promotion in a forum thread gets spotted, downvoted and buried, which means the page a model would have cited now says something you would rather it did not.
The uncomfortable read on this is that source authority in your category was decided by other people, over years, while you were buying keywords. You are auditing a result, not designing one.
Each engine reads a different shelf
The second finding breaks most of the reporting I see. Gemini, ChatGPT, Perplexity and Google AI Overviews pull from noticeably different citation mixes for the same question.
Track them as one combined total and those differences cancel out. You get a single number, it drifts a little each month, and it tells you nothing about which source to work on.
I keep opening dashboards that report AI mentions as one figure. That is the AI-era version of reporting traffic without a channel split, and our industry spent fifteen years learning why that was useless.
A source that carries you inside Perplexity can be invisible inside Gemini. If your buyers mostly use one assistant, an average across four is a worse input to a decision than the one number that actually applies.
Splitting by engine also changes what counts as progress. Gaining three citations in the engine your market ignores while losing one in the engine it uses is a loss reported as a win.
The audit is smaller than you think
Here is the version that fits inside a normal working week.
Write down 10 to 20 questions your buyers actually ask. Not keywords. The sentences someone types when they are three weeks from a decision and do not yet know your brand exists.
Run each one through the assistants your market uses. Record every source cited, and which engine cited it. A spreadsheet is enough for the first pass.
Then sort that list into three piles: pages you own, pages you can influence, pages you cannot touch. The proportions are the finding. In most audits I have seen, the first pile is far smaller than the executive team assumes.
Run the same questions again with competitors in mind, and note where a rival appears as a source and you do not. Open those pages and read them properly.
The gap is usually concrete rather than mysterious. A comparison table you never published, or a specification page carrying real numbers instead of adjectives.
Sometimes it is simply a review profile with 9 reviews on it while theirs shows 140.
What to do with the pile you cannot own
Owned pages are the easy half. Structured product data, pricing, specifications, original measurements. That work is well understood, even when it gets skipped for something more visible.
The influenced pile is where the interesting decisions sit. A review profile moves when you ask customers at the right moment, which is usually the week after a support problem was solved well, not the week after invoicing.
A community thread moves when a real person from your team answers a real question under their own name, repeatedly, without a link in every post.
Both of those are slow, and neither fits a quarterly campaign calendar. That is exactly why they hold. The sources carrying the most weight in AI answers are the ones hardest to manufacture in a sprint.
The third pile, the pages you cannot touch, still earns its line in the report. Knowing that 40 percent of the sources in your category are beyond your reach is a budget argument, not an admission of failure.
I wrote earlier about how brands lose at the synthesis step rather than at the mention, and about how citation rates vary by industry no matter what you do. A source audit is where both of those stop being observations and turn into a task list with owners on it.
Run it once a quarter. The set moves more slowly than rankings do, which is the good news. This is not a treadmill, it is a map that updates at a sane pace.
If you want it measured continuously instead of by hand, GEOflux.ai (geoflux.ai) is built specifically for this. It maps not just whether your brand is mentioned, but why, and what to do about it.
Your rankings tell you where you sit in a list. The source audit tells you who is writing the answer instead of you.