AI Battlecards: Generating Them Without Making Things Up
Evidence In, Citations Out, and a Verification Pass Before Anyone Uses the Card
Short answer
Generate an AI battlecard by supplying the evidence rather than asking the model what it knows. Paste the competitor's pages, ads, reviews and your deal notes, require a source ID on every line, allow "not in evidence" as an answer, then verify prices, dates and quotes before publishing the card internally.
1. Why do AI-generated battlecards make things up?
Because a battlecard is made of specific, current facts about one company, and a model asked from memory will fill gaps with plausible guesses rather than admit it does not know. Prices, plan limits and recent changes are where the guesses land.
Ask a chat model to "write a battlecard for Veltrix" and it will produce something that looks right: tidy headings, confident strengths and weaknesses, a pricing line. Some of it may be accurate. The problem is that nothing on the card tells you which parts, and the parts most likely to be wrong are exactly the ones a seller will quote.
OpenAI researchers have argued that language models hallucinate partly because training and evaluation reward guessing over admitting uncertainty [s2]. For a battlecard, that means the default behavior is to supply a price, a feature or a weakness rather than leave the line blank. The workflow below is built to reverse that default.
The lines that go wrong most often in an ungrounded AI battlecard:
- Pricing and plan limits, which change often and are frequently out of date in anything the model learned from.
- Recent launches, repositioning and new ad angles, which may postdate the model entirely.
- Weaknesses stated as fact without a source, which are the lines most likely to be challenged by a prospect.
- Customer quotes and review statistics, which a model can produce in the right style without any real source behind them.
2. What is a reliable workflow for AI battlecards?
Five steps: collect the evidence, label each piece with an ID, generate the card with a citation on every line, verify the risky lines, and have a human edit and date it. The model does the drafting; the evidence and the human do the trusting.
- Collect the evidence. Gather the competitor's homepage and pricing page as text, their recent ad copy, a few months of their public reviews, and your own notes from deals where they came up. Paste text, not links, so the model reads what you read.
- Label every piece with an ID and a date. Prefix each item with a short ID and the date captured, such as [P1 pricing page, 2026-09-28] or [R14 review, 2026-09-03]. The IDs are what the citations will point to, and the dates are what make the card checkable later.
- Generate with citations required. Use a prompt that fixes the card structure, requires an evidence ID after every line, and tells the model to write "not in evidence" when it cannot support a line. The generation prompt below does this.
- Run a verification pass. In a separate prompt, ask the model to check each line against the cited evidence and remove anything the evidence does not support. Then check prices, plan limits and dates yourself against the live source.
- Edit, date and own it. A named person reads the card, rewrites anything that sounds generic, adds the objection wording sellers actually hear, and puts a last-verified date on each block. Only then does it go to the team.
This mirrors the techniques Anthropic recommends for reducing hallucinations: explicitly allowing the model to say it does not know, grounding answers in direct quotes, requiring citations for each claim, and restricting the model to the documents provided [s1].
3. What prompt should I use to generate the card?
One that pastes the labeled evidence, fixes the card sections, requires an evidence ID after every line, forbids outside knowledge, and makes "not in evidence" an acceptable answer. Copy the block below and replace the bracketed parts.
You are drafting an internal sales battlecard for [YOUR COMPANY] about the competitor [COMPETITOR]. Use only the evidence between the EVIDENCE tags. Do not use anything you know from elsewhere. Write the card with these sections, in this order: Their positioning (their own words), Where they win, Where they lose, Objections they create and our response, Questions a prospect can ask them, Pricing. Rules: after every line, add the evidence ID it comes from in square brackets. If a section cannot be supported by the evidence, write "not in evidence" for that section instead of filling it. Quote their positioning word for word. Include at least two genuine strengths in Where they win. Do not invent customer quotes, numbers or dates. Keep the whole card under 350 words. EVIDENCE: [PASTE LABELED EVIDENCE HERE] /EVIDENCE. Our product, for the response lines: [PASTE YOUR ONE-PAGE SUMMARY AND PRICING].
Three details in that prompt do most of the work. "Use only the evidence" closes off memory as a source. The citation after every line makes each claim traceable. "Not in evidence" gives the model a legitimate alternative to guessing, which is the single instruction most battlecard prompts leave out.
Why require two genuine strengths?
Because a model asked to help you win will lean toward flattering you, and a card with no competitor strengths fails the first time a prospect names one. Requiring strengths, with citations, forces the model to find the evidence that the competitor is good at something, which is exactly what a seller needs to be ready for.
4. How do I get the model to check its own card?
Run a second, separate prompt that gives it the card and the same evidence, and asks it to find a supporting quote for each line and delete any line it cannot support. Then check the high-risk lines yourself.
Below are a draft battlecard and the evidence it was written from. For each line of the battlecard, find a direct quote in the evidence that supports it and write the quote under the line. If you cannot find a supporting quote, mark the line REMOVE and explain in one sentence what is missing. Also flag any line containing a price, a plan limit, a date or a number, whether or not it is supported, with CHECK LIVE. Do not rewrite the card. BATTLECARD: [PASTE DRAFT] /BATTLECARD. EVIDENCE: [PASTE THE SAME LABELED EVIDENCE] /EVIDENCE.
The CHECK LIVE flag exists because a line can be perfectly supported by the evidence and still be wrong today, if the evidence is old. Prices and plan limits get checked against the competitor's live pricing page every time, regardless of what the model says.
| Line type | Check against | Why |
|---|---|---|
| Price or plan limit | Their live pricing page | Changes often and is quoted on calls |
| Recent launch or change | Their announcement, feature page or changelog | May postdate your evidence |
| Weakness stated as fact | The cited reviews or deal notes | The most likely line to be challenged |
| Customer quote | The original review or source | Must be verbatim and attributable |
| Positioning quote | Their current homepage | Homepages change without notice |
5. What should AI never write on a battlecard?
Anything a prospect could check and find false: prices from memory, invented customer quotes, unsourced claims about the competitor's quality, and statements about what they "are planning". Those lines stay human-written or come out.
- Prices or limits not taken from a dated capture of their pricing page.
- Customer quotes that are not copied verbatim from a real review or conversation.
- Claims about their roadmap, finances or intentions, which no public evidence supports.
- Characterizations like "unreliable" or "poor support" without a cited pattern of reviews behind them.
- Objection wording invented by the model. Real objections come from your sellers, in the buyer's words.
The objection block is the one to watch most closely. A model can write very plausible objections, and plausible is the problem: a card full of objections nobody actually raises trains sellers to prepare for the wrong conversation.
6. How do I keep AI-generated battlecards current?
Regenerate from fresh evidence when the competitor changes something, not on a whim, and compare the new draft with the old card rather than replacing it blindly. The diff shows what the model thinks changed, and you decide.
The advantage of an AI workflow is that regenerating a card is cheap. The risk is that each regeneration can quietly drop a hard-won line or introduce a new unsupported one. Keep the previous card, generate the new one from updated evidence, and review the differences.
The trigger for regeneration is the same as for any battlecard update: a pricing or plan change, new positioning, a new ad angle, a launch, a new cluster of complaints or a new objection in your deals. The when to update battlecards guide maps each to the block it affects.
Key Takeaways
Supply the evidence, never ask from memory
Paste their pages, ads, reviews and your deal notes. A model asked what it knows will guess at prices and recent changes.
A citation on every line
Label evidence with IDs and dates, and require an ID after each line so every claim is traceable.
Make "not in evidence" acceptable
Giving the model a legitimate alternative to guessing is the most effective single instruction.
Verify in a separate pass
Have the model find a supporting quote for each line, then check prices, limits and dates against the live source yourself.
Objections come from sellers
Plausible invented objections prepare sellers for the wrong conversation. Use the buyer wording your team hears.
Regenerate on triggers and review the diff
Cheap regeneration is only safe if you compare the new card with the old one before replacing it.
Frequently Asked Questions
Sources
- Reduce hallucinations Anthropic, Claude documentation, September 2026.Recommends allowing uncertainty, grounding in direct quotes, citing sources for claims and restricting to provided documents.
- Why Language Models Hallucinate Kalai, Nachum, Vempala and Zhang (OpenAI and Georgia Tech), arXiv, September 2025.Research paper arguing that training and evaluation reward guessing over acknowledging uncertainty.
Explore More
Related analyses, benchmarks, and industry insights
Related Guides
Glossary Terms
Turn reading into a reaction
Oppira watches your competitors, keeps your playbook current, and drafts the response. See your market clearly by tomorrow.