ChatGPT Prompts for Ruthless Competitor Analysis

Ask ChatGPT to “analyze my competitor Acme” and it will deliver three confident pages about a company it is largely imagining - plausible strengths, invented weaknesses, a positioning statement Acme never wrote. Competitor analysis is the prompt category where the model’s fabrication habit does the most damage, because the output looks like research and nobody fact-checks a SWOT.

The fix is the same division of labour I use everywhere in this prompt library: the model never supplies facts about the competitor. You paste the evidence - their reviews, their pricing page, their ads, their job posts, all public - and the model does what it’s genuinely good at: extraction, comparison and pattern-finding across more text than you’d read carefully yourself.

It’s worth doing systematically, because competition is no longer an occasional event. In Crayon’s 2026 State of Competitive Intelligence report, seven in ten teams say at least half of their sales opportunities are now competitive, and 57.5% report more of their deals are competitive than a year ago. The same report says 49.6% of teams saw their win rate against competitors increase over the past year against 6.3% who saw it fall - which tells me the teams doing this work deliberately are collecting from the ones who aren’t.

Gather this before you prompt

Every prompt below runs on pasted evidence, and all of it is public. My standard collection pass per competitor, maybe forty minutes:

• Reviews: their G2, Capterra or Trustpilot page - the most recent 20 to 30, plus the most critical ones. Reviews are customers doing your analysis for you.
• Their own words: homepage, pricing page, two or three product pages, latest changelog or release notes.
• Their ads: Meta’s Ad Library and Google’s Ads Transparency Center show any advertiser’s live creatives, free. Copy the text of five to ten.
• Their hiring: current job postings, especially in product, sales and marketing.
• Your own pipeline: notes from recent competitive wins and losses, anonymised.

One rule sits above all of it: public information only. If a prompt needs something you couldn’t show a journalist - a leaked deck, a private roadmap - the problem isn’t the prompt. Date-stamp everything you collect too, because competitive intel expires fast and an analysis run on last spring’s pricing page is worse than none.

1. The review-mining prompt

You are a competitive researcher. Below are [N] verbatim reviews
of [competitor], from [source], collected on [date]. [paste,
numbered]

Extract, citing review numbers for every item:
1. What customers praise most, ranked by frequency.
2. Complaints, each with the trigger that caused it.
3. Switching signals: mentions of tools they came from or
   considered leaving for - quote them exactly.
4. The vocabulary customers use for the problem this product
   solves, where it differs from the vendor's own language.

Rules: only the pasted reviews. Thin evidence = say "thin
evidence". No overall verdict - extraction only.

Why it works: this is the same evidence-mining discipline as my buyer persona prompts, pointed outward - and citing review numbers keeps every claim traceable. Section 3 is the money section: people explaining why they switched to your competitor are handing you your churn risks, and people threatening to leave are drafting your comparison page.

2. The feature-matrix prompt

Here is our pricing/feature page: [paste]. Here is [competitor]'s,
copied on [date]: [paste].

Build a comparison matrix: rows are capabilities either product
claims, columns are the two products. Cell values: YES (quote the
claim), PAID TIER (name it), NO CLAIM, or UNKNOWN.

Rules: a capability their page does not mention is UNKNOWN, not
NO - absence of evidence is not absence of the feature. Do not
infer capabilities from marketing adjectives. Flag rows where the
two pages describe similar things in incomparable ways.

Why it works: the UNKNOWN value is the whole prompt. Models fill matrices the way nature fills vacuums, and a battlecard claiming a competitor “lacks” something their page just doesn’t mention is how sales reps get embarrassed on live calls. I always keep the incomparable-rows flag too - “unlimited workspaces” versus “unlimited projects” is a positioning choice, not a fact you can tabulate.

3. The positioning-teardown prompt

Here is [competitor]'s homepage and main product page copy,
copied on [date]: [paste].

Answer from this text only:
1. Who they are selling to - role, company size - as evidenced by
   the words they chose, with quotes.
2. The one problem they claim to solve, in their words.
3. What they emphasise that we do not, and vice versa. Here is
   our homepage for comparison: [paste].
4. What they conspicuously do NOT say - missing proof, absent
   audiences, unclaimed use cases. Label this section SPECULATIVE.

Why it works: copy is strategy leaking in public. Question 4 finds the gaps - the audience their words ignore is often the one you can own - but it’s also where the model tips from reading into guessing, which is why the section carries a label. I treat everything under SPECULATIVE as a hypothesis to test, never as a finding to repeat in a deck.

4. The ad-angle prompt

Below are [N] of [competitor]'s current ads, copied from
[Meta Ad Library / Google Ads Transparency Center] on [date]:
[paste ad text, numbered].

Analyze: (1) the angles in play - problem-led, proof-led, offer-
led, fear-led - with counts; (2) the offers and CTAs; (3) which
audience each angle implies; (4) angles from our own playbook
that are absent from theirs.

Rules: analyze only the pasted text. Ad volume is not evidence
of ad performance - do not describe any angle as "working".

Why it works: the ad libraries show you what a competitor is testing, and the “not evidence of performance” rule keeps the analysis honest about what running ads can and can’t tell you - a live ad is a bet, not a result. The absent-angles list in point 4 feeds straight into my ad copy prompts as candidate angles for your next batch.

5. The hiring-signals prompt

Here are [competitor]'s current job postings, collected on
[date]: [paste titles + key responsibilities].

What do these hires suggest about their next two quarters -
new markets, new product areas, a motion change (PLG, enterprise,
partnerships)? For each inference: the postings supporting it and
a confidence level (strong / weak). Everything in this output is
SPECULATIVE by nature - say what would confirm or kill each
inference.

Why it works: job posts are the most forward-looking public signal a company emits - they spend money on them, and they describe the future, not the present. But inference from hiring is inherently soft, which is why I bake the epistemics into the prompt: confidence levels plus a confirm-or-kill condition turn gossip into a watchlist.

6. The win/loss mining prompt

Below are anonymised notes from [N] recent competitive deals -
wins and losses against [competitor]: [paste].

Extract: (1) stated reasons we won, ranked by frequency;
(2) stated reasons we lost; (3) the objections prospects raised
that came from the competitor's talking points, quoted;
(4) the moment in the deal where competitive pressure appeared.

Rules: reasons must be stated in the notes, not inferred from
outcomes. Where a note gives no reason, count it as "reason
unknown" - do not backfill.

Why it works: your pipeline is the one competitive dataset nobody else has, and it corrects everything the public sources exaggerate. The “no backfill” rule matters because models pattern-match losses to plausible causes, and a fabricated loss reason is worse than an unknown one - you’ll fix the wrong thing. Run this quarterly and diff the results, the way I recommend refreshing personas against new evidence.

7. The battlecard prompt

Using only the outputs of the prompts above [paste them], build a
one-page battlecard for sales:

- Three things they beat us on, stated plainly, with our best
  honest response to each.
- Three things we beat them on, each backed by evidence a rep
  can cite - a review quote, a matrix row, a win reason.
- Their likely talking points against us (from the win/loss
  objections), each with a reply and the evidence behind it.
- Landmines: questions a rep can ask that expose their gaps -
  only from UNKNOWN or complaint evidence, never invented.
- Date of the underlying evidence, stated at the top.

Rule: any claim without a traceable source from the pasted
outputs gets cut, not softened.

Why it works: the battlecard is where analysis either becomes revenue or becomes shelfware, and this is exactly where the market is racing: in the Crayon report, 80% of teams now use AI to generate sales-facing competitive content, up from 25% just two years earlier. I insist on the “honest response” framing for their advantages - a battlecard that pretends the competitor has no wins trains reps to lose credibility in the exact moment they need it. And the evidence date at the top enforces the expiry logic: Crayon’s data shows teams sharing intel weekly or faster achieve revenue impact at 79% versus 41% for monthly-or-slower, which is an argument for a light process you rerun over a heavy one you archive.

The cadence that makes it compound

My working rhythm, for a marketer covering competition alongside everything else: the full collection-and-prompts pass quarterly per major competitor, the review-mining and win/loss prompts monthly (they move fastest), and the battlecard updated whenever either of those shifts a claim. The first pass takes a couple of hours. Repeat passes are faster because you’re diffing against a baseline, and the diff itself is the insight - what changed in their reviews, their ads, their hiring since last time is more actionable than any single snapshot. And if competitive intelligence is the center of your job rather than one duty among many, the product marketer playbook covers the whole seat.

Where the outputs flow: gaps from the positioning teardown become content brief angles, absent ad angles feed the next ad copy batch, and the customer vocabulary from review mining sharpens personas and copy alike. Competitor analysis that ends life as a slide deck changed nothing - the point of doing it with prompts is that the outputs are already in the shape the rest of the workflow eats. For where this sits in the bigger practice, start with the complete guide to AI marketing.

FAQ

Can ChatGPT just browse my competitor’s site itself? Models with browsing can fetch pages, but I still paste evidence deliberately. You control what was collected and when, the analysis stays reproducible, and you avoid the failure mode where the model quietly mixes fetched fact with trained-in folklore about a company. The forty-minute collection pass is the audit trail.

Is this legal and ethical? Everything above uses public information - reviews, published pages, ad libraries, job boards - which is ordinary market research. The line I hold: nothing obtained through misrepresentation, nothing confidential, and nothing scraped in violation of a platform’s terms. If you wouldn’t put the source in the footnote of a board slide, don’t put it in the prompt.

How many competitors should I run this on? Deeply, two or three - the ones actually showing up in your deals, which the win/loss prompt will tell you. I’d rather have a current, evidence-dense card on the competitor in half my pipeline than shallow folders on ten companies my prospects never mention.