Competitive Intelligence When the Shortlist Comes Out of an Assistant: Keep the Log, Not the Verdict
Your rival list is produced by the questions you ask, so the asset worth maintaining is a dated log of which domains held the slots for which question, on which engine.
Published by AI Knows Us · Updated 16 October 2026
Competitive intelligence changes shape when the shortlist a buyer works from arrives inside an answer rather than from a results page. What you can record is narrow and reliable: for a written question, asked on a named engine, on a date, which companies were named in the answer and which domains were credited as sources. That is a log, and the log is the asset. A verdict is not available, because the same question asked twice in one day can name different companies, and because the rival list you see is an artefact of the questions you chose. Freeze the questions first, then count.
The classic side of this work is covered separately in what to watch when you monitor competitors in search. This page is the answer layer beside it: what can be counted, and what the count is worth.
The unit of observation, and why it has to be fixed before anything else
An answer is not one source. A run dated 16 October 2026, putting 210 buying questions per market to ChatGPT and to Gemini across three markets, recorded 20,143 citations over 1,260 readings and 4,947 distinct domains. Deduping to one domain per reading, that is an average of 15.99 credited sources behind every answer. A single answer therefore stands on roughly sixteen domains, and a rival can sit in the prose, in the sources, in both, or in neither.
Four events, logged in four columns, never in one:
- Named in the answer text. The company appears in the words a buyer reads.
- Credited as a source. A page on that company's domain is one of the sources the answer was built from.
- Named through somebody else's page. The name sits inside a list or review page that was credited, so the name travelled and the credit went elsewhere.
- Put forward as a choice. The prose tells the reader to consider them, which is a judgement about language and needs a written rule before scoring starts.
Collapse those four into one number called visibility and the log stops being usable, because they move for different reasons.
Your rival list is produced by your question set
This is the finding that matters most before anybody builds a tracker, and it came from running two question sets eleven days apart in design.
The narrower run, 102 questions per market on 15 October 2026 and 597 usable readings of 612 attempted, had four specialist tools among its ten most credited domains: tryprofound.com in 61 readings, otterly.ai in 58, depra.ai in 53 and peec.ai in 49. The wider run, 210 questions per market on 16 October 2026, had none of that kind in its top ten. Its leaders were developers.google.com in 446 readings, support.google.com in 328, developers.google.cn in 325, blog.google in 309 and help.openai.com in 269, which is a set of platforms answering questions about themselves.
Same category, same two engines, same week. Widening the questions replaced the visible competitor set with a documentation layer. Two consequences follow directly.
- Freeze the question file with a version number and a date and leave it alone between rounds. A rival list that moves because you added questions is not intelligence.
- Split the file in two: questions a platform will always answer, such as how an engine picks sources, and questions a supplier can answer. Judge rivals on the second group only, because a dozen questions from the first group will flatten every round you run.
The five fields that make a log worth keeping
| Field | Why it earns its column |
|---|---|
| Question text and file version | A question retyped differently is a new question, and its row is not comparable to last month's |
| Engine, named individually | A blended figure hides which engine a rival owns |
| Every credited domain, copied exactly | The row that tells you who holds the slot. Keep the raw string, normalise in a second column |
| Which of the four events applied | Named, credited, named through a third party, put forward. Four columns, not one |
| Date, time and location setting | Answers differ by all three, and a log without them cannot be re-run by anybody |
Weekly suits most categories and monthly suits slow ones. What matters more than frequency is that the question file does not change, because the comparison between rounds is the whole value.
The intelligence is on the rival's page, not in the answer's prose
Sorted by who owned the credited page, 82.0% of those 20,143 citations sat on a vendor's or an agency's own website. Official platform documentation took 12.3%, news or large publishers 3.2%, social or community 1.9%, and review directories 0.7%.
Read that as a map of where to look. The slots not held by platform documentation are held by somebody's own pages, which are readable, dateable and comparable, and the credited page is the real intelligence product. Keep a short dossier per rival, updated whenever a round credits them:
- The exact URL that was credited, and the date.
- The checkable fact on it: a coverage list, a dated price, a named integration, a stated limit, a sample size.
- What changed on that page since your last round. A rival who added a dated price is telling you what they intend to win.
- Whether you publish the equivalent fact, and where.
That last line is the only one that produces work. Most of what separates a credited page from an ignored one is a document rather than a product, and a document can be written this month. Hold the 0.7% figure in mind when somebody proposes a year of directory profiles as the competitive response.
Four things the log cannot tell you
- Market share. Nobody knows how often real buyers ask your questions, or in what words, so no honest denominator exists for a share claim. The log describes your question file.
- Why. An engine's account of its own choice is a statement rather than a mechanism. Keep it in a quotes column and out of the numbers.
- A per-industry rate. Those runs asked generic questions by market, so they hold no rate for software, healthcare, legal or any other sector. A sector rate comes from a question file written in that sector's words.
- One blended score. One of the two engines credited our own domain in 0 of 921 readings across both runs while the other credited it in 6 of 1,260 in the wider one. An average of two engines that differ that much describes neither.
A one-off count of how often a named rival is recommended ahead of you is a different exercise with its own protocol, written up in a repeatable brand recommendation test. The log is the ongoing version, and it is the cheaper habit.
Questions about competitive intelligence from AI answers
How many questions should the file hold?
Between twenty and one hundred, in the words a buyer types, with no company name in any of them. Below twenty, the round-to-round movement is larger than anything you are trying to see. Above one hundred, the file becomes expensive to re-run by hand, and re-running it unchanged is the point.
Can the answer tell me who my competitors are?
It tells you who holds the slots for your questions, which is a more useful list and often a different one. Expect three kinds of name: real rivals, list pages and directories sitting between you and the buyer, and platforms answering questions about themselves. The second kind is usually addressable by appearing on the page rather than by outranking it.
How do I handle a rival with a name close to another company?
Write the matching rule down first: the legal name, the trading name, the product name and the domain, with a note on which near-matches are excluded. A keyword match will credit a similarly named company to your rival, and that error is invisible afterwards.
Does classic keyword and ranking research still feed this?
Yes, and it is where the question file usually comes from. The queries where you sit just below the first page are the contests worth writing questions about, and finding every keyword a website ranks for covers how to pull a rival list.
How often does a rival list actually change?
Less often than the prose does. Individual answers vary within a day, which is why repeats belong in the method, while the set of credited domains across a frozen file tends to move over months. Report the domain set as the trend and keep any single answer as evidence.
What a rival's credited page is really telling you
Run the file once and the most uncomfortable row is rarely a rival name. It is a question where sixteen sources were credited and none was yours, on a subject you have published about for years. That row is not a judgement about your product, because no engine has tested anybody's product. It is a statement about which documents were readable, specific and dated enough to stand behind a sentence.
Our own row in that log is published rather than hidden, and it is a poor one: 6 credits in 1,260 readings, 0.5%, on the date above. The method sits in what AI assistants actually cite across 597 readings. AI Knows Us is paid to keep that log and to write the pages that argue for a different name in it. Starting the log by hand costs an afternoon, and the first round is worth running before anybody sells you a dashboard, because then you know what it should be counting.