Five sites read, then eight: what the source count beside an answer is

The number of pages an app read before answering is a field, and it is the field most people throw away. In the scan published on our proof page, run 2026-10-16 on ChatGPT and Perplexity, the four published readings carry source counts of 5, 5, 6 and 8. Keeping that figure beside each answer changes what the rest of the record can tell you, because an answer built from 5 pages and an answer built from 8 are different kinds of evidence even when the text looks similar.

The four readings, with the count attached

Question and app	Sites read	Outcome
Pricing question, ChatGPT	5	Named the company first, with four other names
Pricing question, Perplexity	5	Named the company first, with three other names
Coverage question, ChatGPT	6	No business named at all
Coverage question, Perplexity	8	Named two publishers, not the company

The pattern in that table is worth reading slowly. The two readings that produced a confident first name read the fewest pages. The reading that produced no name at all read more. The reading that read the most pages produced the most cautious naming, two established publishers rather than a ranked list.

What the count is a reading of

It is a reading of how much work the answer needed, which is a proxy for how much agreement the app found. A question with a well published answer gets settled quickly from a few pages. A question where the sources disagree, or where no page makes the comparison being asked for, takes more reading and produces a hedged or nameless answer.

Treat that as a signal about the question rather than a measurement of the category. Four readings cannot establish a general relationship between source count and answer type. What they can do is tell you which of your own questions are cheap for an app to answer and which are not, and the expensive ones are where the publishing opportunity is.

There is a second use for the field that has nothing to do with interpretation. It tells you whether the app searched at all. An answer produced with no sources behind it came out of what the model already holds rather than out of the current web, and a reading like that describes a memory rather than a state of the world. Those two kinds of answer should not sit in the same count, and the only way to separate them later is to have written the source count down at the time.

Why the count has to be recorded at the time

Because it is not recoverable afterwards. Re-run the question a week later and you get a different set of pages and a different count, so a reading with no source count is a reading missing a field that can never be filled in. The same is true of the source list itself, which is the part most worth keeping: the domains named instead of yours are a shortlist somebody else's buyers are being handed.

Our own wider readings are built on exactly this discipline. On 15 October 2026, across 597 usable readings on ChatGPT and Gemini, 10,652 citations were recorded across 2,708 distinct domains, every source URL written down. Without the per answer source record there would be no source mix to report, and the source mix is the finding: 85.6% of those citations went to the site of a company or agency selling the thing being recommended, and 0.8% went to review directories.

The five fields worth keeping for every answer

• The exact question, word for word, with no company name in it.
• The app and the date, because both change the answer.
• Whether your business appeared at all, and separately whether it appeared in the sentence doing the recommending rather than in the background.
• Every domain the answer read or cited, as domains rather than page titles, so they can be counted.
• The number of sources read, which is the field this page is about and the one that gets dropped first.

A log with those five fields can answer questions a single scan cannot, and the structure of keeping one is set out in keeping the log rather than the verdict. The verdict is what people remember and the log is what can be re-read.

Two things the source count will not do

It will not tell you how competitive a market is, because a thinly published question in a crowded market can be settled in a few pages. And it will not substitute for reading the sources themselves: the count tells you how many, and the list tells you who, and only the second one tells you what to write. Across four readings on one day the count is a prompt to look at the list rather than a finding in its own right.

Questions about how many sources an AI app reads

Is a higher source count better for me?

Neither is better. A low count on a question you win is a comfortable position, and a high count on a question you lose is where there is room to publish something that settles it.

Where do I find the number?

In the interface, where the sources behind an answer are listed, and it needs writing down at the time because the next reading will differ. Record the domains at the same time, since both come from the same place.

Does a cited source mean the answer used it?

Being listed and being leaned on are different things, and worth separating in the log. Being named and being linked are two outcomes, and a log that stores them in one column cannot tell you which one you have.

Should I report the source count to a client?

Yes, as a column rather than a headline. It is what makes a one line answer auditable later, and it costs nothing to keep.

Why does a company selling this care about one field?

Because we are paid to take these readings and write the pages that follow them, and the pages are chosen from the source lists. Without the count and the list, a reading is an anecdote with a date on it.

One field, and the record it makes possible

A scan that reports whether you were named is a verdict. A scan that reports whether you were named, what it read to decide, and how many pages that took, is a record. The first can be repeated and compared only on its headline. The second can be re-read months later by somebody who was not there, which is the test any measurement has to pass before a decision gets made on it. Four readings on one day is a small record, and it is a record rather than a verdict because that one field was kept.