AI competitor analysis: prompts, sources and pitfalls

AI competitor analysis means using an assistant such as ChatGPT, Claude or Perplexity to sort and summarise what you know about your competitors. It is very good at the sorting and summarising. It is unreliable at anything it has to remember or guess: a rival's current price, when something launched, whether they are running ads this week.

The working rule for competitor analysis with AI: you supply the evidence, the assistant does the analysis, and you check the output against the evidence. Below are the prompts that work, how to feed in real sources, and where the answers go wrong.

What AI does well and badly in competitor analysis

Where the assistants make things up

OpenAI's own research says it plainly: language models hallucinate "because standard training and evaluation procedures reward guessing over acknowledging uncertainty" (Why language models hallucinate, September 2025). The same piece explains that "arbitrary low-frequency facts", its example is a pet's birthday, "cannot be predicted from patterns alone". Most competitor facts are exactly that kind of fact.

Three places it bites:

Prices. Ask for a rival's pricing and you will often get a confident table. It may come from a page that has since changed, from a third-party listicle, or from nowhere. Plan names that no longer exist are the giveaway.

Dates. When a product launched, when a feature shipped, when prices last changed. Low-frequency facts, the kind that get guessed.

Ad activity. "They are running Google Ads on these terms" or "their Meta ads focus on discounts", with no source. An assistant cannot know what ads a competitor is running this week unless it reads the ad libraries or you paste the ads in. Treat any unsourced claim about current ads as invented.

Turning web search on shifts the risk rather than removing it. The answer cites pages, but a page may be old, or may not say what the answer claims. OpenAI's own help page tells users to "open a cited source to check that it supports the answer, review when it was published or updated, and use an authoritative source when accuracy matters" (Searching the web with ChatGPT). That advice applies to every assistant. Some sites also block the crawlers assistants use to fetch pages, so a competitor's own site can be missing from the answer altogether; which AI crawlers to allow explains the difference between those crawlers.

How to feed it real sources

The fix is to do the collecting yourself, or with a tool, and hand the assistant evidence with dates on it.

1. Collect. For each competitor: the text of their pricing page, their homepage hero and main landing pages, the URL list from their sitemap, ad text and start dates from the ad libraries, 30 to 50 recent reviews, and any Reddit threads about them.
2. Label every item with the competitor, the source URL and the date you collected it.
3. Paste it in or upload it as files. Keep one project or conversation per analysis so the context stays together.
4. Restrict the sources. Tell it to use only what you supplied. If you do want it to search, point it at specific sites: ChatGPT's deep research, for example, can be limited to sites you list (OpenAI's help page).
5. Demand evidence for every claim, then check a sample yourself.

The ad libraries guide covers collecting ads, which is the input people most often skip and the one assistants most often invent.

ChatGPT prompts for competitor analysis that work

Each prompt assumes you have already pasted or uploaded the evidence. They work the same in Claude or Perplexity. Change the brackets.

1. The ground rules. Start every session with this.

You are helping me analyse competitors. Use only the material I provide in this
conversation. For every claim, say which item it came from. If the material does
not answer a question, say "not in the sources" rather than guessing. Do not
state prices, dates, ad activity, traffic or revenue unless they appear in the
material.

2. Positioning comparison.

Below are the homepage headline, subheading and first three proof points for
[Competitor A], [Competitor B], [Competitor C] and us. For each: what category
do they put themselves in, who is the stated or implied audience, and what is
the main promise? Then tell me which two are hardest to tell apart, and quote
the words that make them similar.

3. Review mining. The reading method behind it is in finding positioning in competitors' reviews.

Here are [40] reviews of [Competitor], with star ratings and dates. Group the
complaints into themes. For each theme, give the number of reviews that mention
it and two short quotes. Do the same for praise. Leave out reviews about a
single one-off incident.

4. Pricing comparison from pages you collected.

Here is the text of each competitor's pricing page, with the date I copied it.
Build a table: plan name, monthly price, annual price if shown, main limits, and
what the cheapest plan leaves out. Mark anything ambiguous instead of
interpreting it. Add a column with the date each page was collected.

5. Ad messages.

Here is the text of every active ad I found for [Competitor] in the Meta Ad
Library and the Google Ads Transparency Center, with each ad's start date.
Group the ads by promise and by offer. Which messages appear most often, and
which ad has been running longest? List the landing page URLs in order of how
many ads point to each.

6. Sitemap analysis.

Here is a list of URLs from [Competitor]'s sitemap with their last modified
dates. Group them by page type (blog, location, product, comparison,
integration, other). Count each group, and list every URL modified in the last
90 days. What does the pattern suggest they are investing in?

7. The red team.

Act as head of marketing at [Competitor]. Using only the evidence about us
below, write the three arguments you would use to win a customer who is
comparing us. Then tell me which of those arguments our evidence can answer and
which it cannot.

8. The check.

List every factual claim in your last answer as a table: the claim, the item it
came from, and a direct quote from that item. Mark any claim that has no
direct quote.

Prompt 8 is the one people skip, and the one that catches the most mistakes.

A one-hour AI-assisted competitor analysis

For three competitors, once the evidence is collected:

• 10 minutes: the ground rules and the positioning material, then prompt 2.
• 15 minutes: each competitor's reviews through prompt 3.
• 10 minutes: pricing pages through prompt 4.
• 10 minutes: ads and sitemaps through prompts 5 and 6.
• 5 minutes: prompt 7, against your own evidence.
• 10 minutes: prompt 8, then spot-check five claims against the original pages.

The output is a draft, not a decision. Turning it into three actions with owners and dates is still the human part, described in how to do competitor research.

Mistakes that waste the hour

Asking an open question. "Analyse my competitors" with no material gets you a plausible essay built from whatever the model half-remembers. Every useful prompt above starts with "Here is".

Letting it choose the competitors. Assistants name the best-known brands in a category, not the businesses you actually lose to. Bring your own list, built from sales calls and search results as in finding your real competitors.

Mixing dates. A pricing page from March next to one from September produces a comparison that was never true at any single moment. Collect everything in the same week.

Trusting the counts. "12 of 40 reviews mention support" is easy to check and often slightly wrong. Recount the biggest theme yourself before you quote it.

Pasting half a page. Pricing tables often hide limits in footnotes or tooltips. If the cheapest plan's limits are missing from what you pasted, the comparison will be too generous.

Choosing an assistant for competitive analysis

The method matters more than the model. The practical differences:

• Perplexity shows its sources with every answer, which makes checking quick.
• ChatGPT and Claude both search the web when search is switched on, both take uploaded files, and both handle long pasted material well. Claude's help centre explains turning web search on.
• Public competitor material is safe to paste. Be more careful with your own confidential material, such as customer lists or deal notes, and check your plan's data settings first.

Dedicated AI tools for competitor analysis exist too; the roundup of AI competitor analysis tools compares them.

Where Figo fits

Ours, so judge accordingly. The hard part of AI competitor analysis is the first step, collecting current evidence, and that is the part Figo does. Each week it records new ads from the Google, Meta and TikTok ad libraries, new and changed pages, pricing page changes, rankings, social posts, Reddit threads and review ratings for each competitor. Claude writes the Monday briefing from that stored data and nothing else, and Ask Figo answers questions from the same data.

Figo does not track what ChatGPT, Claude, Gemini or Perplexity say about your competitors. If that is what you are after, it is the wrong tool.