Citation probability

What does a citation probability estimate look like?

The strip at the top of this page is one, from our own workspace. In the 30 days to 4 October 2026, ChatGPT linked attensira.com in 154 of 2,349 answers to the 61 questions we track, and Perplexity in 85 of 2,175, read over the Attensira MCP on 4 October. Those are observed rates: 6.56% and 3.91%.

The probability behind each is not that exact number. It sits somewhere in a range set by how many answers we read. With samples in the thousands the ranges are narrow, about 5.6% to 7.6% for ChatGPT and 3.2% to 4.8% for Perplexity (95% Wilson score intervals). They do not overlap, so "ChatGPT links us more often than Perplexity on this question set" is a claim the sample can carry. Had we read 30 answers per engine instead, the same rates would give ranges so wide that nobody could say which engine cites us more.

The overall figure, 239 of 4,524 answers or 5.28%, is real but says little about any single question. Citation probability belongs to a question, not to a site, and it is spread very unevenly across the questions we track.

What is citation probability?

It is the chance that an AI engine attaches a link to your domain when it answers a specific question. It belongs to a combination of things, not to your site:

• the exact wording of the question;
• the engine and the model behind it that day;
• the engine's search index and what it fetched;
• the country and sometimes the account asking.

Change any one and the probability changes. That is why a single "citation probability" figure is meaningless without the question, the engine and the sample. Two of our questions that differ by a few words, both about what AI search visibility costs, show how far apart the probabilities can sit.

How is it measured?

By sampling. Ask the question, read the answer, record whether your domain was linked; repeat. The observed citation rate, answers citing you divided by answers read, is the estimate. Attensira's citation rate is exactly that, returned with the number of answers behind it.

Two rules from Attensira's reading-the-numbers guide keep the estimate honest:

• Below 20 answers, show the count, not a percentage. "4 of 12" carries its own uncertainty. The same figure written as a percentage hides it.
• A change must clear the noise. Two rates are compared with a two-proportion test at 95% confidence. A movement that does not clear it is reported as "nothing we can prove yet".

What raises citation probability?

The levers are upstream of the answer:

1. Access. The engine's search agent can fetch the page. A page that is never fetched has a citation probability of zero for every question. See robots.txt.
2. A passage that answers the exact question. Retrieval works on passages, so the page needs a section that answers this question on its own. See chunking.
3. Presence on preferred sources. Each engine leans on certain domains for certain questions. If they cite a review site, being on it raises the chance your brand rides along, even when your own domain is not linked.

Citation is also not the same as being named. On "AI search visibility starting price", ChatGPT linked attensira.com in 42 answers, yet Attensira was named in 0 of the 109 answers to that question across both engines. See AI brand mentions.

What are the objections to estimating citation probability?

"Give me a predicted probability for my site." No one can predict it honestly for questions nobody asked. Measure the questions that matter to you.

"Our rate went from three in twelve to five in twelve." With 12 answers each, that difference is well inside the noise.

"One engine is enough to estimate it." On the starting-price question, ChatGPT linked us in 42 of 57 answers and Perplexity in 0 of 52. Same question, same month.

"More pages means more probability." Only if the new pages answer questions better than the old ones. Near-duplicates split what you had.

How does Attensira estimate and raise it?

Attensira reads your questions on each engine every day, so the sample behind each rate keeps growing, and reports every rate with its count. It flags thin samples and tests every change before calling it real. The crew then works the levers above, shipping each change after your approval and reading the rate before and after, with its interval. Citation tracking shows each rate with its count, and our methodology sets out when a sample is too thin to quote.