Why does AI competitor research so often go wrong?
AI competitor research goes wrong when you ask a tool what it knows about a company instead of giving it sources to read, because it will produce a plausible table from memory, mixing old facts, other companies' features and pure invention. The table looks finished, which is what makes it dangerous.
Imagine a fictional 30-person software company, Quillpost, which makes invoicing software for freelancers. Its product manager, Amara, receives this message from the chief executive on a Monday: 'Can you look into what Northbill and Paperkite are doing? Heard Northbill does AI invoice chasing now. Are we cheaper? Need something for the board on Thursday.' Northbill and Paperkite are fictional competitors for this lesson.
The request mixes a rumour, a pricing question and a deadline, and it does not say what decision the board is making. That is normal. Your first job is to turn it into research questions.
How do you do competitor research step by step?
Do competitor research by starting from the decision, writing a short list of questions, choosing the sources that can answer them, and filling a comparison table where each cell holds a finding, a link and the date you checked. AI speeds up the reading and extraction; you choose the questions and verify the evidence.
1. Name the decision
Amara confirms the board is deciding whether to prioritise automated payment reminders next quarter. That narrows the research to reminders, pricing for comparable plans and target customers.
2. Write five to eight questions
'Does Northbill offer automated reminders? Does it describe AI in them? Which plans include them? Who do they say the product is for?'
3. Rank your sources
Official product pages, help documentation, pricing pages and changelogs first; public filings such as the SEC's EDGAR system in the US or the Companies House register in the UK for company facts; reviews and forums last, labelled as opinions.
4. Give AI the sources
Paste text from the pages you collected, or use a research tool that browses and returns links, and tell it to answer only from those sources.
5. Fill the table with evidence
Each cell gets the finding, a short quote, the URL and the check date. If a source is silent, the cell says 'not stated on official pages', which is different from 'no'.
6. Write the 'so what'
Three lines on what the findings mean for the decision, separate from the facts.
What does a weak and an improved competitor table look like?
Weak prompt: 'Compare Quillpost with Northbill and Paperkite.' Typical result: a tidy eight-row table claiming Northbill has 'AI-powered smart reminders', 'over 50,000 customers' and a starter plan at a specific monthly price, and that Paperkite 'lacks automation'. None of it has a source. The customer number and price could be invented, and 'lacks automation' may simply mean the tool found nothing.
Improved prompt: 'Using only the source extracts below, fill this table. Rows: automated reminders, AI described in reminders, plans that include reminders, stated target customer. Columns: Northbill, Paperkite. In each cell give the finding, a quote of no more than 20 words and the source URL. If the sources do not answer, write not stated. Do not use your own knowledge.'
The improved table shows that Northbill's help centre describes scheduled reminder emails and its product page uses the phrase 'smart reminders' without explaining what is automated, so the AI claim stays unconfirmed. Paperkite's reminders appear in its help documentation for paid plans only. Customer numbers are 'not stated', because neither company publishes one. Amara can now tell the board exactly what is known, what is rumour and what to check next.
How do you check competitor research before you share it?
Check competitor research by opening every link in the table, confirming the quote is on the page and still current, and reading each cell as if the competitor's own product manager were reviewing it. If you would be embarrassed to show them the cell, it is not ready.
Click every link
AI research tools can attach a real link to a claim the page does not make. Confirm the quote is there.
Date everything
Pricing and features change often. Put the check date in the table so readers know how fresh it is.
Separate silence from absence
'Not stated' means you found no evidence. Only write 'no' when a source says so.
What are the common mistakes in AI competitor research?
The biggest mistake is trusting a table built from the model's memory, which mixes outdated and invented facts with real ones. Others: researching everything instead of the decision, comparing your internal roadmap with a competitor's live product, treating marketing phrases as features, and quoting prices without the date. Stay ethical too: use public information, do not misrepresent who you are to get a demo or a quote, and never use confidential material from a competitor's former staff.
Where does this lesson sit in the Tosheo.ai catalogue?
Competitor research is a Practitioner lesson in the AI for Product category, and marketing learners often meet it too because campaign positioning depends on the same evidence. The full category is on the AI training for product teams page. The hands-on task gives you a fictional brief and source extracts, and the rubric checks your questions, your evidence in each cell and how you handle what the sources do not say.
Sources and scope
GOV.UK: get information about a company from Companies House, checked 19 October 2026 ↗