Charleston AI search: a different answer every time

Go open Gemini right now and ask it for the best electrician in Mount Pleasant. Read the answer. Then ask the exact same question again.

You will probably get a different electrician.

Not a slightly reordered list. A different business, sourced from different websites. Same phone, same account, same words, sixty seconds apart. I have run this on Charleston queries enough times now that it stopped being surprising and started being the whole point, and this week somebody finally put a number on it.

14,472 citations, and the local pack wins on consistency

Ben Fisher at Steady Demand published a study on August 19 that I think is the most useful piece of local search research of the year so far. His team ran 1,487 local queries through Gemini across all 50 of the largest US metros and 10 service categories. Lawyers. Dentists. Auto repair. Plumbing. HVAC. Electrical. Pest control. Locksmiths. That is not an abstract SaaS dataset. That is a list of my clients.

They logged 14,472 citations, then repeated the identical queries to see whether the answers held.

They did not.

• Ask Gemini the same local question back to back, and the cited domains overlap 46.3% of the time
• Wait 3.5 hours and ask again, and overlap drops to 41%
• Wait until the next day and it falls to 26.5%
• Gemini recommended the same top business on repeat about 7% of the time

Now the control. Fisher ran the same repeat test against Google's traditional local pack. It returned the same top listing 90.2% of the time.

That gap is the story. The map pack is a ranking. You earn a position, you hold it, you can watch it move. Whatever Gemini is doing is not that.

Why it happens, in one sentence

Fisher calls it grounding drift, and the cause is further upstream than most people assume.

When you ask Gemini a question, it does not search the web with your words. It writes its own internal search queries first, then reads what those return, then answers you. And it writes different internal queries every single time. Across repeated identical calls, Fisher measured only about 5% overlap in the search terms Gemini generated for itself.

So the model is not being flaky about which business it likes. It is reading a different slice of the internet on every run, because it went looking with different words. Byte-identical input, different search, different sources, different answer.

(This is also why two owners in the same Charleston zip code can both "check ChatGPT" on the same afternoon and come away with completely opposite conclusions about who is winning. They both ran a sample size of one.)

At this point a lot of owners are thinking: so AI search does not matter

That is not what I am saying, and I want to be careful here because the lazy version of this post gets used as an excuse to ignore the whole channel.

AI search matters. What I am saying is that an AI mention is not a rank, and you should stop managing it like one.

Concretely, three things follow from that:

Stop buying the single-check screenshot. If an agency sends you a screenshot proving you are cited in Gemini, that screenshot is worth roughly what one coin flip is worth. Fisher's own framing is that a single "am I cited" test is closer to a coin flip than to evidence. Ask for the same query run several times, on different days, with a couple of phrasings. If they cannot produce that, they are selling you a lucky roll.

Stop panicking about a single bad result. You asked, a competitor came up, and now you want to burn the whole strategy down. Ask again tomorrow. Then ask three more times. Judge the pattern, never the instance.

Stop assuming one engine tells you about the others. Gemini and ChatGPT cited the same domains only 8% of the time on identical questions, and named the same top business 4.2% of the time. Being the answer in one tells you almost nothing about the other. I covered a related version of this problem when I wrote about how Charleston businesses winning the map pack can be invisible in AI answers, but that was about a visibility gap between two surfaces. This is worse. This is instability inside a single surface.

The Reddit lesson, which happened this month

If you want proof that building on somebody else's platform is a rented position, look at what just happened to Reddit.

Promptwatch tracked Reddit's share of ChatGPT Search citations at a steady 3.83% from July 18 through August 7. By August 14 it was under 1%, averaging 0.52% through August 17. An 86.4% relative collapse in about four days. Search Engine Land reported it on August 20, Forbes picked it up the same day, and the leading explanation is that Reddit blocked its whole domain in robots.txt.

I wrote in July that Reddit was quietly deciding a lot of AI answers and that Charleston service businesses should get into those conversations. I still think showing up in real local threads is good for you. But the citation channel that made it urgent evaporated inside a week, through a decision nobody in our industry got a vote on.

That is the whole risk profile of platform-dependent tactics in one chart.

The one source that held

Here is the part of Fisher's data I keep coming back to, and it is genuinely good news.

Across all 14,472 citations, nearly 60% of what Gemini cited was the business's own website. Reddit was second at 13.7%. Every local-service directory combined, Angi and Thumbtack and HomeAdvisor and the rest, accounted for 10.3%.

Your site beat the entire directory category by roughly six to one.

ChatGPT weights things differently, citing business websites only 15.9% of the time, which is exactly why you cannot optimize for one engine and call it done. But note what does not appear anywhere in the drift data: your own site does not get deprecated by a robots.txt change at a company you do not control, and it does not vanish when a licensing deal expires. It is the one input to these systems that you actually own. That is not a coincidence, and it is why I keep pushing Charleston SEO work toward the site itself rather than toward whatever citation surface is fashionable this quarter.

What to actually do this week

• Run the drift test on yourself. Pick your money query, something like "best HVAC repair in West Ashley." Run it in Gemini three times today and once tomorrow. Write down who gets named. That five-minute exercise is a better AI visibility audit than most paid dashboards.
• Fix the pages Gemini is already reading. Services, service areas, pricing, and a real about page with named humans on it. Sixty percent of citations went to business websites, so the question is whether yours can answer the question when it gets read.
• Do not consolidate onto one off-site channel. Directories, local press, community threads, and your Google Business Profile are all backup routes. Reddit just demonstrated why you want more than one.
• Change how you measure. Twice a month, same queries, same phrasings, logged in a spreadsheet. Trend line, not screenshot.

The businesses that come out of the next two years fine are not the ones who cracked the prompt. They are the ones whose own website was so clearly the best answer about plumbing in Summerville that it did not matter which slice of the web the model happened to read that morning.

That is boring. It is also the only part of this that no one can take away from you on a Tuesday.

Run your three queries today, and write down what you see.