Request an audit

Field notes

Ask an AI Twice, Get Two Different Shops

AI can answer from memory or search the web first. In one real audit, the same shop came out very differently each way. Here's why that gap matters.

When a driver asks an AI assistant where to take their car, it can answer in one of two ways.

It can answer from memory, using what it learned in training, without looking anything up.

Or it can search the web first. It looks at a handful of pages first, then writes its answer from what it found that day.

Same assistant. Same question. The answers about your shop can be very different.

What One Real Audit Showed

I'll use our sample report, because you can open it and check every number. It's a real audit from October, with the shop's name and town changed. The shop is called Maple Street Auto.

We asked ChatGPT, Claude, Gemini and Perplexity the same 50 questions drivers ask when they need a shop, three times each. Perplexity always searches, so it only answered one way.

From memory, Gemini recommended the shop for 4% of the questions. After searching the web, 32%.

ChatGPT went from 2% to 10%. Claude went from 0% to 8%.

Same shop. Same questions. Same day. Nothing about the shop changed between those answers. Only where the AI looked.

That's one shop, not a law of nature. But it's a big gap, and it points at something every shop owner should know.

Memory Is Old News

An assistant answering from memory is working from whatever it absorbed when it was trained. That can be months or years old.

A shop that opened recently, changed its name, or started doing diesel work last spring may simply not be in there.

When the assistant searches, it looks at what's online now: your website, your listings, the pages that mention you. If those pages say clearly who you are and what you fix, the assistant has something to work with.

Your Customer Doesn't Pick the Mode

Drivers don't choose memory or search on purpose. Whether an assistant looks things up depends on the assistant, the question and its settings. Perplexity, for one, always searches, so it has no memory answer to measure.

So the same driver asking the same question could get your name one day and someone else's the next.

Neither of you would know why.

Why We Never Blend the Two

A Spilld report measures memory answers and search answers separately and shows them side by side, assistant by assistant.

One blended number would hide the part that matters. A shop that shows up when AI searches but vanishes from memory is in a different spot than a shop that's invisible either way. The first gets found when AI goes looking. The second doesn't get found even then, and that's a different problem.

It's also why we compare each assistant against itself. Perplexity has no memory answer, so lumping every assistant together on each side would compare two different groups.

One Search Is a Guess

If you've typed your town and "auto repair" into ChatGPT once, you saw one answer, in one mode, on one day. It's a start. It isn't a measurement.

Even in the same mode, answers move around. That's why we ask every question three times, from memory and with web search wherever the assistant allows it.

Want to see memory and search side by side? The sample report is at spilld.ai. The memory-versus-search table is on page 5.

Sources

  1. Spilld, Public sample report: Maple Street Auto (shop name and town changed)

See what AI says about a shop like yours.