AI Competitive Intelligence: What It Cannot Find

AI can tell you what a competitor has published. It cannot tell you what they charge, what they are building next, which of their accounts is at risk, or why you lost a deal to them. I tested that in the week commencing 7 September 2026. Neither Crayon nor Klue publishes a price. Three third party pages priced Crayon anyway, from $12,700 to over $100,000 a year, and none of them agreed with the others.

A disclosure before anything else. ORRJO sells competitive intelligence, and it sells against Crayon and Klue by name through a comparison page for each. I used them as the test subjects because they are the two platforms whose material I have read most closely, which makes me a fair reader of it and an interested one. Every figure below comes from their own site or from a page I opened and have linked. Where a number flatters them, it stays in.

Can ChatGPT do competitor research?

It can retrieve and summarise what has already been published about a company. It cannot verify any of it, and it's bad at telling you when it doesn't know.

The best public measurement of that second problem is the Tow Center for Digital Journalism study "AI Search Has a Citation Problem", by Klaudia Jaźwińska and Aisvarya Chandrasekar, published in Columbia Journalism Review on 6 March 2025. They ran 1,600 queries across eight generative search tools. Each query handed the tool a direct excerpt from a news article and asked it to name the headline, publisher, date and URL. Every question had one checkable right answer. The tools got it wrong more than 60% of the time. Perplexity was the strongest at 37% wrong. Grok 3 was wrong 94% of the time.

The finding I care about more is the refusal behaviour. ChatGPT Search identified 134 articles incorrectly, signalled any lack of confidence 15 times out of 200 responses, and never once declined to answer.

That study tested news attribution rather than competitor research, and I'm not going to pretend the headline rate transfers. What transfers is the shape of the failure. A tool that won't say "I don't know" about a fact with a single right answer won't say it about your rival's renewal terms either. It will produce a number.

The test: what does Crayon cost?

Price is the first question on every competitive brief I've been handed, and it is the question the public record handles worst. So that is the one I ran.

The question was: what does Crayon's competitive intelligence platform cost per year? I put it to public search in the week commencing 7 September 2026, then opened every source that answered instead of reading a summary of them. This is what each page actually says.

Two errors in that table will not survive summarisation, and both of them matter.

The Rival Radar page contradicts itself. It gives one range in its comparison section and a wider one in its own table and FAQ, and an engine reading the page has no way to know which the author meant. Separately, the two cheapest alternatives in the category are the two sources quoting the highest ceilings for Crayon. Read any price alongside the business model of whoever published it. A summary strips that context out by design, which is how a ceiling quoted by a $49 a month rival ends up sitting next to a median drawn from 93 recorded purchases, the two of them looking equally like facts.

Ask an engine what Crayon costs and you'll get a confident single number, probably close to $30,000, because that is the figure that recurs. It may well be right. Vendr's median comes from 93 recorded purchases and is the only figure in the set with a stated method behind it. The problem is that the number reaches you with the method, the date, the sample size and the publisher's interest all removed.

The same question for Klue returns the same pattern. Vendr gives an average and median of $30,000 a year and a range of $16,000 to $60,000, based on 106 purchases, also last updated February 2026. klue.com/pricing publishes no figure and asks for a demo.

What is the best AI for competitive analysis?

That is the wrong question, and the way it is wrong is the interesting part.

In the same week I searched "best competitive intelligence software 2026" and, rather than reading what the nine first page results said, I looked at who published them. Five are the blogs of companies selling software in or next to the category. One is a directory that sells placement. One is an agency. Two describe themselves as media, and one of those takes affiliate commission.

Some specifics, because the pattern is easier to argue with than a summary of it.

• Contify's page is titled "Best Competitor Analysis Tools & Software for 2026" and ranks Contify first. Contify sells what it calls an "AI-native market and competitive intelligence platform". The page carries no disclosure that the publisher is also the winner.
• Autobound places its own Signal Data API at number 15 in its own list of 15, and does not say that it is Autobound's.
• Improvado does disclose, inside the guide: "Improvado is our platform, and it is scored on the same criteria as every tool here". That is the standard the rest should be held to.
• Software Advice discloses the most plainly of all, because its model requires it: "Vendors who have paid for placement have a 'Visit Website' button, whereas unpaid vendors have a 'Learn More' button."

Unkover, one of the two media pages and the one taking affiliate links, ran a wider count than mine and reports that "of the top 19 results currently ranking for 'competitive intelligence tools' on Google, 14 are published by vendors promoting their own software". I haven't reproduced that count and am reporting it as theirs. My smaller one points the same way.

So the answer to "what is the best AI for competitive analysis" is assembled almost entirely out of documents written by people who sell the answer. None of that is sinister. It is what happens when a category's buyers ask a question that only its vendors are motivated to answer in public. The practical consequence is that the ranking you are reading is a map of who published, and it should be read as one.

What can AI not find out about a competitor?

Anything that was never published. That reads as obvious and it gets forgotten constantly in practice, because a fluent answer feels like a researched one.

Below is the grid I work from. Thirteen questions taken from a real competitive brief, marked against what public retrieval could actually produce about Crayon and Klue in the week commencing 7 September 2026. The value of a grid like this sits in the blanks rather than the ticks.

Rows 8 to 13 decide deals, and no tool at any price reaches them, because the information is not in public. Row 13 is the one worth arguing about, because it is the only blank with a known method attached. Win-loss interviews reach it, and they are slow and manual by design. The slowness is where the answer comes from.

Rows 1 to 6 are the ones people mean when they say AI has changed competitive intelligence, and they have a point. Retrieving and cross-checking six directory profiles used to be an afternoon and is now a minute. What changed is the cost of the easy half. The hard half costs what it always did.

Is using AI for competitive intelligence legal?

Reading what a competitor has published is not a legal problem, and the reason sits in the definition of a trade secret rather than in anything to do with AI.

The Trade Secrets (Enforcement, etc.) Regulations 2018 set out, in regulation 2, what has to be true before information counts as a trade secret at all:

• it is secret, meaning it is not generally known among, or readily accessible to, people in the circles that normally deal with that kind of information;
• it has commercial value because it is secret; and
• the person lawfully in control of it has taken reasonable steps to keep it secret.

A page a company publishes on its own website fails the first of those on the day it goes live. Nothing on the open web is a trade secret, and a machine reading it faster does not make it one.

The exposure sits in how non-public information is obtained. Two provisions do most of the work in the UK.

• Computer Misuse Act 1990, section 1. Causing a computer to perform any function with intent to secure unauthorised access to programs or data, knowing at the time that the access is unauthorised. On indictment that carries up to two years. Scraping content that sits behind a login you were never given is the everyday version.
• Trade Secrets Regulations 2018. These give a trade secret holder a route to remedies against an "infringer", defined as a person who has unlawfully acquired, used or disclosed a trade secret. The regulations do not define unlawful in the abstract, which is exactly the point. It turns on how you got hold of it.

Then there is the part no statute covers tidily, which is misrepresentation. Phoning a competitor's support line as a prospect, or their sales team as a student, is where competitive intelligence programmes usually get into trouble, and it has nothing to do with AI. The trade body for the profession, SCIP, publishes a code of ethics that addresses it. Its site did not resolve when I tried to open it during this test, so I'm not going to quote it from somebody else's summary.

The short version. If the information is on the open web, a machine reading it is legally uninteresting. What somebody does when the machine comes back empty is where the trouble starts.

Why the engine's answer about your competitor is your competitor's marketing

This is the part that should change where money goes.

If an engine can only build an answer out of documents that exist, then the answer a buyer receives about your category is decided by who published what. Crayon's price is a vacuum, so two firms selling at $8 and $49 a month have filled it with numbers of their choosing, and those numbers now sit one search away from any buyer who asks. Klue put a customer count in a press release in 2021, and five years on that is still the figure sitting closest to hand.

Turn the same mechanism around and it becomes your problem rather than theirs. When a buyer asks an engine who the good platforms or agencies are in your category, the answer is assembled from your competitors' pages and their review profiles. If you haven't published, you are not in the answer, and being better will not put you there. That is the whole argument for answer engine optimisation, made from the buyer's side of the table rather than the marketer's.

ORRJO's website audit is £750 and runs live buyer prompts against the engines to see who gets named when somebody asks for a recommendation in your category. The monitor that re-runs the same prompts every month is £149 a month. Neither replaces research. They tell you what the machine currently believes about you, which is a different question from what is true. The wider picture is in the state of GTM research in 2026.

How do you run this test on your own competitors?

An afternoon and a spreadsheet.

1. Write the brief first, as questions. Thirteen or so, like the grid above, each phrased so it has a checkable answer. Do this before you open any tool, or the tool decides what you asked.
2. Answer each one from the competitor's own site. Pricing page, newsroom, careers page, customer page. Record the URL and the date you looked. That is your control.
3. Then ask the engines, and open every source they cite. Read the source page itself rather than anybody's summary of it, and note who owns the domain and whether they sell something competing.
4. Mark the disagreements instead of resolving them. Where two credible sources give different numbers, the gap is the finding. Averaging destroys it.
5. List what stayed blank. That list is your research agenda, and it is the only part of this exercise worth paying anybody for.

The blank list is where ORRJO Intelligence starts, from £2,500 a month, and it is why the work is done by a named analyst rather than a dashboard. The method, including how findings reach a sales team without becoming another document nobody opens, is in our guide to competitive intelligence for B2B.

A dashboard will fill row 7 with $30,000 and move on. An analyst writes "not published" in row 7, and then goes and asks the buyers who chose them.