AI search volatility: how to measure through the noise

AI search volatility: how to measure through the noise

LA
Jedi Solomon
··5 min read

AI search volatility is high for a single answer and much lower for a pattern. Ask ChatGPT, Gemini or Perplexity the same buyer question three times and you can get three different lists of businesses. Run the question on a schedule, count how many runs name you, and the number settles enough to act on.

Why AI answers change between runs

Several things move at once. The model varies its wording each time, so a list of names can reorder or drop a business. Retrieval also varies. When an engine searches the web before answering, it draws from a wide pool. Ahrefs (March 2026) found that 38% of AI Overview citations come from the top 10 results and 31% from positions 11 to 100. A large share of sources sit well off the first page, and which of them gets pulled is not fixed.

Context moves too: location, conversation history and model updates. Local answers add another layer. Gemini grounds them on Google Maps, while ChatGPT and Perplexity use Bing Places and Yelp. A fix to one listing can show up in one engine and nowhere else.

What one run can and cannot tell you

A run that names you does not prove you are cited. A run that leaves you out does not prove you lost ground. Each reading has a job:

ReadingWhat it tells youUse it for
One runThat the answer is possibleSpot checks, seeing what an engine says about you
Three runs, same dayWhether you appear most of the timeScoring a single prompt
Same panel, every monthDirectionReporting

We define "cited" as named or linked in at least 2 of 3 runs on the measurement day. A business that appears in one run out of three is borderline, and the rule keeps it from being counted as a win.

A method that survives AI search volatility

  1. Fix a panel of about 25 buyer prompts and freeze the wording. Changing a word changes the answer. The audit page shows how we set that panel.
  2. Run every prompt three times per engine, on the same day, in a fresh session with no saved memory.
  3. Record for each run whether you were named, linked or absent. Record which competitors appeared.
  4. Score a prompt as cited only at 2 of 3 runs. Report the share of the panel you are cited on.
  5. Repeat on the same day each month with the same engines and the same wording.
  6. Log every change you make, with its date: a page published or a listing fixed. Without the log, you cannot connect a movement to a cause.

Weekly spot checks are fine for catching a broken page or a wrong claim. Draw conclusions from the monthly panel only.

Report each engine on its own line

Engines differ more than runs do. The SOCi Local Visibility Index 2026 found ChatGPT recommends 1.2% of multi-location businesses, Gemini 11% and Perplexity 7.4%. A blended average across the three hides that spread and makes every month look like a muddle. Give each engine its own line. A business can be strong in Gemini because of a clean Maps profile and invisible in ChatGPT because its Bing Places listing is thin.

Good AI citation tracking also keeps a competitor column. If you slip on a prompt and the same rival shows up in every run, something real changed. If a different name appears each time, you are looking at normal drift.

Traffic is the steadier series

Citations are the leading signal, but visits are summed over many answers, so individual swings average out. At JCS Fashions, a 4,878-product Shopify apparel store in Milpitas, California, sessions from AI assistants went from 20 in January 2026 to 569 in October. Orders from AI assistants were 5 in October. That count is too small to read month-to-month movements into, so we treat orders as a slow signal and lean on citations and sessions. The JCS Fashions story has the full set of numbers, including the season effect on total sales.

Pair the panel with your analytics. The panel says whether engines name you. Sessions from assistants say whether anyone follows through. When both rise over three or four months, AI visibility measurement is telling you the same thing from two directions. When they split, check the landing page the engines send people to.

What to do this week

  • Write down 25 questions a buyer would ask before choosing a business like yours.
  • Run each one three times in ChatGPT, Gemini and Perplexity, and put the results in a spreadsheet with the date.
  • Mark each prompt cited or not cited using the 2 of 3 rule.
  • Book the same exercise for the same day next month.

That is 225 runs, a long afternoon by hand. If you want a baseline first, the free AI visibility scan shows where you stand today, and you can compare your own panel against it.