Measuring AI visibility: a weekly prompt × engine matrix

Short answer. Write 5–20 prompts in your customers' words, ask each engine the same prompt weekly, and record two booleans per answer: was the brand named, was the site cited as a source. Aggregate into mention rate and citation rate per engine and per prompt, keep the full answer and the searches the model ran, and compare week over week. One run is noise; the trend over 4–8 weeks is the signal.

The two numbers

For every answer an assistant gives, record:

• Mentioned: does the answer name your brand (any of your brand terms)?
• Cited: does the answer link to a page on your domain as a source?

Everything else, position in a list, sentiment, length, is a refinement. These two are enough to see whether the AI layer knows you and whether it sends people to you.

Aggregate them into a matrix: prompts as rows, engines as columns, a mention rate and a citation rate in each cell, plus a row and column total. seops’s demo workspace for the made-up game VoidX looks like this after 28 days:

Engine	Mention	Cited
ChatGPT	66.7%	40.0%
Gemini	63.3%	40.0%
Perplexity	80.0%	80.0%
Claude	70.0%	50.0%

The pattern is typical: Perplexity cites almost whenever it mentions; ChatGPT and Gemini mention more than they cite.

Writing prompts

Write them the way a customer types, before they know your name:

• “Non-violent space games for kids on iPhone”
• “What are good space games I can play offline on my phone?”

Then a few navigational ones where the brand is in the prompt:

• “Is VoidX free? Does it have in-app purchases?”

Five to twenty prompts is enough. More prompts don’t improve the signal; they dilute your attention. Rotate one or two a quarter as your product changes.

Sampling

• Same prompt, same engine, once a week, same day and hour.
• Keep the full answer text and, where the engine exposes it, the searches it ran. The searches tell you which queries the model used to find sources; those are keywords worth ranking for in Bing and Google.
• Keep the cited URLs, yours and others’. The “cited instead of you” list is your competitor set in the AI layer, and it rarely matches your Google competitor set.

Reading the trend

One run is noise. Compare 7, 28 and 90-day windows. A change of 10 points week to week means little; a change of 10 points over 8 weeks is real. seops shows both rates with the point change over the previous window for exactly this reason.

Two things move citation rate in practice:

1. Being in the Bing index with a crawlable, fast page. ChatGPT search and Copilot read Bing. See the IndexNow post.
2. A page that answers the prompt in its first paragraph, with the entity names the model is looking for. Long introductions lose to short direct answers.

What not to do

• Don’t count a mention in a list of ten as success. Record it, but weight citation.
• Don’t optimise for a single engine; the four disagree, and the mix of your customers’ assistants changes.
• Don’t claim causation from one week.

Running it

In seops: Growth › AI visibility › Add prompts, set brand terms in project settings, and the Monday 07:00 run does the rest. Each cell opens the full answer. Your assistant can read the same data through the get_ai_visibility and get_ai_answer MCP tools.

Questions