One Question, Five Engines, Five Different Companies

01Ask five assistants, get five shortlists

Put the same category question to five assistants, something like which tool is best for sending invoices, and the lists that come back rarely match. One names a company the others ignore. Two agree on a leader and disagree on everything underneath. A brand that tops one answer is missing entirely from another. It is tempting to call that noise. It is mostly structure.

Five engines, at least four different pools

Each assistant builds its answer from a different place. ChatGPT mixes what it remembers from training with OpenAI's own search index and live fetches. Perplexity searches an index its own crawler builds, weighted toward recent pages. Claude's web search runs through Brave, which Anthropic lists as its search provider, before fetching the pages it picks. Copilot is grounded in Bing. Google's AI Mode fans a question out across Google Search. Different pools produce different candidates before any model has chosen a word. There is no such thing as being cited. There is being cited by one of them.

Even two surfaces from one company disagree more than you would guess. Google's own AI Mode and AI Overviews share only 13.7% of their citations, so five engines from five companies drifting apart is the expected result rather than a malfunction.

02Run it the same way every time

Set up so your history cannot help you

Use a fresh session on every engine: signed out, or in a temporary chat, with memory and personalisation switched off. Ask a category question that does not contain your name, word it identically everywhere, and do not follow up, because a follow up tells the model what you wanted to hear. If the assistant remembers you asked yesterday, the test is measuring your history, not your reputation.

Three runs each, fifteen in total

Answers vary from one run to the next, so a single run proves almost nothing about anyone. Ask each engine three times in separate sessions and write down every company named, in the order it appears, plus the one sentence each engine uses to describe you when it does. Fifteen runs take about twenty minutes. It is the same discipline as the answer audit, turned sideways: one question across many engines instead of many questions to one. One run is an anecdote. Three runs is a pattern you can defend.

Write down the description as carefully as the name. Being named with an old price, a feature you removed or the wrong category is a different failure from not being named at all, and it points somewhere else: to the facts you publish and repeat, rather than to whether any index has found you.

03What the pattern is telling you

Named by one, missing from the rest

Presence on a single engine points at the pool you are strong in, and absence elsewhere points at the pools that have not met you yet. Named by Copilot and nobody else usually means Bing knows you well. Named by Perplexity alone suggests recent pages or community mentions are carrying you. Present in ChatGPT only with browsing off means memory has you while current search does not. Each gap has a different fix, and none of them is a generic push for more content. Absence on one engine is a map of which index has not met you yet.

Named by none, and the same rivals everywhere

The harder result is being absent from all fifteen runs while the same two or three competitors appear in most of them. That is not a technical problem you can find in a crawl. It means the sources every engine draws from already agree on the category without you, and the fix lives on pages other people write. When five engines agree on your competitors, the market has already written the list.

Keep the sheet and repeat it monthly with the same wording. Your share is simply your appearances out of fifteen, set beside each rival's, and the movement between months is worth more than any single score. Numbers from one day can mislead. A trend recorded the same way twelve times is hard to argue with.

One honest limit. Fifteen runs is a small sample, engines change their behaviour without announcing it, and one category question does not describe a whole market. Treat the sheet as a smoke detector rather than a census: very good at telling you something is wrong, much weaker at telling you precisely how wrong.

Tomorrow: which of those engines ever sends a person back, and which only take.