How AI Assistants Decide Which Businesses to Recommend

Many business owners are puzzled when they ask ChatGPT on a Monday which shops do the best auto glass in Spokane, then ask the same question on Tuesday and get a different set of names. That happens because AI assistants are designed to give varied answers when they recommend businesses. It isn't a bug, and it doesn't make tracking your AI visibility a waste of time.

What it does tell you is that your spot in any single answer means very little. The figure worth following is how often your business comes up when the same question is put to the assistant many times over, and a local business can usually move that share more easily than a national brand can.

A Hundred People Asking the Same Question Get a Hundred Lists

Imagine 100 people in Spokane and Spokane Valley need a new windshield this week, and each asks an AI assistant for help, wording the request differently. One person asks for the best auto glass in Spokane, while another wants a chip fixed on a Subaru near the South Hill. A third person has been talking to ChatGPT about her commute for months, so it already remembers where she works.

They get 100 different lists, with different businesses, a different order, and different sources cited underneath.

If you own one of those shops, which answer is your ranking? None of them. Each is a snapshot of something that reshuffles every time it's asked.

Researchers Logged 2,961 Runs, and the Lists Rarely Matched

In January 2026, Rand Fishkin of SparkToro and Patrick O'Donnell of Gumshoe.ai tested this with about 600 volunteers and 2,961 runs across ChatGPT, Claude, and Google's AI. Per SparkToro's write-up, ask ChatGPT or Google's AI the same thing 100 times, and there's less than a 1-in-100 chance any two responses give you the same list of brands. The same list in the same order came back about 1 time in 1,000.

That makes tracking your position in AI answers a measurement of noise. Third today and seventh tomorrow says nothing about your business.

The set of brands each AI drew from stayed fairly stable, though. Same contenders, shuffled differently every time. That stability is why Fishkin backed off his own assumption that tracking AI visibility was useless.

Stop Chasing Rank, Start Raising the Odds

If you ask an AI model like ChatGPT the same question 100 times and record the names, the most common business might show up in 85 of the provided answers. The next spot gets maybe 30, and we’ll say the small long-tail shop receives only two.

No ranked list would give you figures like those. A pool would, though, if every business in it had its own weight and each answer picked out only a few of them, so that asking again brought back a different result.

Nobody controls the draw. What you can change is your weight in the pool, and the share of answers you appear in tells you whether that's working.

Four Layers Decide Who Gets Drawn Every Time Someone Asks

How does AI recommend businesses if there's no fixed ranking to win? There's no evidence of a hidden formula or brand-specific tuning. What looks like an algorithm emerges from four layers working at once, and any one of them can swap a business in or out of a given answer.

1. Training data determines what the model already knows, which is how big national brands show up in answers before anyone searches for anything.

2. Live retrieval splits the question into smaller searches, a process called query fan-out, and pulls in pages from those results.

3. Built-in randomness varies the output, partly by design and partly from tiny rounding differences on the servers running the model.

4. Personalization shifts the answer based on memory, location, and whatever came earlier in the conversation.

Of these four, live retrieval usually decides things for a local business. The person asking for the best auto glass in Spokane won't see it. Still, the assistant may also be running extra searches in the background, for example, on mobile windshield repair, on insurance-approved installers, and on same-day chip repair.

If the question is worded slightly differently, a different group of those sub-searches gets set off, and they bring back other pages and other businesses. A page that ranks for the main query and several sub-queries therefore has more than one route to the same answer. That is generally why covering a topic in depth works better than having a single page perfectly tuned to one query.

Most niche and local businesses get in this way. A model probably learned very little about a three-van glass shop on Division Street during training. It finds that shop live, usually on someone else's page, like a review site or a "best of Spokane" list.

What Other Sites Say About You Outweighs Your Own Website

When Ahrefs looked at 75,000 brands, the number of times a brand was mentioned across the web tracked visibility in AI Overviews much more closely than backlinks did, with a correlation of 0.664 compared with 0.218 (Ahrefs' study of 75,000 brands). The data comes from May 2025 and shows correlation rather than causation, but the gap between those two figures is large enough to take seriously.

Retrieval explains it. An assistant has no reason to take a business at its word, so if your site calls you Spokane's best glass shop and nothing else online agrees, the claim carries almost no weight. When a local roundup, a couple of review sites and a neighborhood forum all name you independently, that's corroboration, and corroboration is what gets a business into the pool.

Most of the evidence that decides whether you're drawn therefore sits on pages you don't own. Your website is still important, but it can't stand alone as a reference.

Does Reddit Help You Show Up in AI Answers?

It can, but the live citations most people chase have proven an unreliable source of visibility. According to Semrush's three-month citation study, nearly 60% of ChatGPT's prompt responses cited Reddit in early August 2025, and by mid-September that figure had dropped to around 10%. A business whose whole AI plan was to get mentioned on Reddit would have seen the channel shrink over a few weeks even though it hadn't changed anything on its end.

The training-data layer is more durable. Whatever communities have written about your business over the years has already become part of what the model knows, and it stays there even when the share of live citations falls.

There is a limit to how far this can be gamed, though. Assistants look for agreement among independent sources, so a few real threads from people who actually hired you count for more than a batch of planted ones. When many posts come from a single source disguised as multiple sources, they usually still register as one voice.

A Smaller Local Pool Means Every Mention Counts for More

In a national category, hundreds of businesses compete for the handful of slots that each answer has room for. A local query such as "best auto glass Spokane" works from a much smaller set, perhaps 10 to 15 businesses, which the assistant puts together from only a few kinds of input, mainly Google Business Profile data, review sites, and a few "best of" lists.

Having so few candidates helps a local business in two ways. Fewer businesses get pulled in, and each new independent mention shifts the weighting much more than it would in a crowded national category. A single honest write-up in a Spokane roundup makes up a real portion of everything the assistant can find about auto glass in the area, whereas for a national brand one more write-up hardly registers.

That's why moving from 20 of 100 answers to 60 of 100 is a realistic target here. Run that math in a national category and see what it costs.

Feed Both Layers with Authority the Internet Can Confirm

At August SEO, that's the job Authority Engineering is built for: real expertise on your own site, plus corroboration elsewhere on the internet that backs up what you say. Each half feeds a different layer.

The on-site half feeds retrieval. Fan-out runs searches on the sub-questions buyers ask, such as whether insurance covers the repair, how long a chip repair takes, or whether you'll come out to an office in Spokane Valley, so content that answers those questions is what gets retrieved. The other half, independent third-party mentions, feeds the pool, since those mentions decide which businesses are eligible to be drawn in the first place. If you build only one, you either get retrieved and then passed over, or you get named with nothing behind the name.

What you should be tracking, then, is a share: out of 100 answers to the questions your buyers really ask, the number that mention your business, along with the sources that led the assistant to you. In a pool as small as Spokane's, you can realistically raise that number, which is more than can be said for chasing the #1 spot in ChatGPT.

A simple count is a good place to begin. Take a handful of questions your customers really ask, run each one through ChatGPT, Claude, and Google's AI several times, and tally how often your business is named and which pages are cited. The cited pages will show you where to start working.