Using AI for Competitor Research: Give Every Number a Trail

When you use AI to build a competitor report, the footnotes are the part to trust least. They look like evidence. Often they point to a real page that never says the thing the sentence claims. The fix is cheap if you build it in from the first row: every number carries its own trail back to the original sentence.

Why can't you treat AI footnotes as sources?

Because a citation being attached is different from a citation supporting the claim. In 2023, Nelson F. Liu, Tianyi Zhang and Percy Liang at Stanford audited four generative search engines: Bing Chat, NeevaAI, perplexity.ai and YouChat. On average, only 51.5% of the generated sentences were fully supported by their citations. Looked at from the other side, only 74.5% of citations actually supported the sentence they were attached to.

Roughly half the sentences had footnotes that didn't hold them up.

In March 2025 the Tow Center for Digital Journalism at Columbia University ran an easier test. Researchers gave eight AI search tools 200 news excerpts each, 1,600 queries in all, and asked them to name the source article's headline, publisher, date and URL. Every excerpt was one where a plain Google search put the original in the top three results. The tools got more than 60% of the queries wrong. The best, Perplexity, was wrong 37% of the time. Grok 3 was wrong 94% of the time. ChatGPT gave 134 incorrect answers and signaled doubt in only 15 of them.

That was an open-book exam. A competitor report asks for headcount, revenue, customer counts and prices scattered across annual filings, press releases and job ads. It is a harder exam.

Write the question and the definition before the AI starts

If you don't define the number, the AI picks a definition for you and doesn't say which one.

Take "how many people does the competitor have in Japan?" Employees only, or contractors and agency staff too? The Japanese subsidiary alone, or every group company in Japan? As of when? Leave those open and you may get the "3,000 employees group-wide" line from the corporate site, which is a different thing from what you wanted to compare.

So the first column of the worksheet is the question itself, written precisely enough that someone else could go and find the same number: "Full-time employees of Company A's Japanese subsidiary, standalone, as of March 31, 2026." Write the denominator too. Revenue per head with employees as the denominator and revenue per head with contractors included can lead you to opposite conclusions.

If the number can't be found, write "not disclosed." I put this in my ground rules for a reason. An empty cell is better than a guess, and what a company chooses not to publish tells you something.

Let the AI find sources. You rank them

The AI is fast at listing candidate sources. A person decides which ones go into the sheet.

The ranking follows the six source levels in the same ground rules. From the top: binding documents the company itself issued, like terms and press releases; audited or statutory records, like Japan's annual securities reports and corporate registrations; job postings; material from agents and partners; signed media reports; interviews. AI research usually surfaces candidates from the first five. Interviews you have to go and do yourself. A number that appears in a media report gets checked against levels one to three before it goes in the sheet. Search snippets and the AI's own summaries don't rank at all.

One more split matters. A number the company states about itself goes in as a claim. "Over 10,000 client companies" on a website is the company's word. It can go in the sheet, labeled as a claim, until an independent source backs it.

Every number should trace, along its own row, back to one sentence in the original

Structure of a competitor-report worksheet, one row per number. Citation support rate: Liu, Zhang and Liang (2023), four generative search engines

1. 1Question and definitionWho, which unit, as of when, and what the denominator is. Written before the AI searches
2. 2Candidate sources: AI lists, you chooseRanked on the six source levels. Media figures checked against higher levels. Self-reported figures marked as claims
3. 3Evidence: URL, location, date retrievedLocation down to page, table and line. Paste the sentence containing the number, word for word
4. 4Calculations on their own rowState which rows the numerator and denominator come from. Match the years
5. 5Conclusions written separatelyCite row numbers only. A judgment with no event behind it is marked "to verify"

A footnote is not proof the sentence exists in the source. If you can't paste the original sentence, the number stays out.

Give every number a URL, a location and a date

A URL alone isn't enough. An annual securities report in Japan can run well over a hundred pages. A link to the whole PDF means whoever checks your work starts at page one. Write the page, the table and the line.

The retrieval date matters too. Company websites change, and investor relations pages get updated. An old number with a date is still a fact: "the corporate site said 3,000 on October 20, 2026." The same number with no date becomes something nobody can explain six months later.

The single most useful habit is this one. Ask the AI to paste the sentence that contains the number, unchanged. Then open the page yourself and search for it. If it can't paste one, or the pasted sentence isn't on the page, the number was probably filled in by the model. My guess is that a good share of the sentences outside that 51.5% would fail right here.

Keep calculated numbers on their own row

Revenue per employee, market share and growth rates are all calculated. None of them appears in the source. Don't mix calculated numbers into the same column as copied ones.

Give each one its own row and say which row the numerator comes from and which the denominator comes from. Whoever checks it can see at a glance whether both sides use the same year and the same definition. Fiscal 2025 revenue divided by April 2026 headcount is an easy mistake to make. The AI does sometimes get division wrong. What worries me more is a denominator quietly swapped, and without the formula written out, nobody can see it.

Write the conclusion apart, citing rows

Separate the conclusions from the numbers. Each judgment carries its row numbers: "Company A's Japan sales team is roughly twice ours (rows 7 and 9)." Anyone who disagrees can go straight to those two rows and argue there.

Some judgments have no event behind them, like "they seem to be shrinking their reseller network." You can write those down. Mark them "to verify" and keep them outside the conclusion. A competitor map is meant to be redrawn, and as I wrote in The two maps, every point on it should be able to say where it came from.

None of this is specific to AI. When research was slow, people took notes as they went because they had time to. AI made the searching fast, and the note-taking is what quietly fell out. It is the same split I described for AI and interview notes: hand over the sorting, keep the moment where you decide something is a fact.

Next time you ask an AI for a competitor table, leave the location column empty and ask it to fill that in. The rows it can't fill are the ones to go and read yourself today.

---