ChatGPT product positioning prompts for marketers

These ChatGPT product positioning prompts give B2B marketers copy-pasteable templates for ICP definition, competitive differentiation, positioning statements, and messaging pillars. They go further than the usual prompt lists by running April Dunford’s and Geoffrey Moore’s methods step by step, chaining outputs across a four-step workflow, and pressure-testing the result against your own win/loss and interview data.

For putting voice rules into a reusable artifact, see how to train an LLM on brand voice. For the operating system these artifacts sit in, see how to build an AI content workflow. Once the statement is approved, AI prompts for LinkedIn posts is the next place those claims get used.

What should you gather before you prompt?

You determine positioning output before you open ChatGPT. When you provide no customer evidence or product proof, ChatGPT tends toward generic positioning, and prompt wording alone does not supply the missing context. The inputs below are the context artifacts you build once and reuse in every session. The prompt is where you invoke them, not where they live.

• Company profile: What the product does, who buys it, what it replaces, how it is priced, and every proof point you can defend (customer outcomes, third-party tests, certifications). Keep it to a page.
• ICP notes: Closed-won accounts from the last one to three years with industry, headcount, contract value, time to close, and renewal status, plus churned accounts with the stated reason. Add the buying roles you see in deals.
• Competitor list from the customer’s side: Dunford’s method starts with the alternatives your best customers would have used if you did not exist, which usually includes a spreadsheet, an intern, or doing nothing. Capture homepage, pricing, and two case-study pages per competitor, dated.
• VoC data: Verbatim language from customer interviews, win/loss calls, call recordings, and reviews. Do not substitute CRM closed-lost fields for this. Clozd’s analysis of buyer interviews, cited in Pragmatic Institute’s 2023 win/loss report, found the CRM closed-lost reason was wrong 85% of the time and the wrong competitor was tagged 65% of the time.
• Voice guidelines: Tone, banned phrases, and a few samples of on-brand copy. You need these for the messaging step, not the positioning step.

Store these in one place every session can reach, whether a shared doc or a ChatGPT Project, and paste them into the context block of each prompt below. Teams that skip this step and type product details from memory into each chat get different positioning every time they ask.

How do you write a positioning prompt that stays adaptable?

Every prompt in this article has the same five parts in the same order, which is what lets you swap in your own product without rewriting the template. OpenAI’s developer-message structure runs identity, instructions, examples, then context, with the context block placed near the end of the prompt and separated by delimiters.

• Role: One line that sets the working lens (competitive analyst, buyer-persona researcher). Keep it relevant to the task and stop there. The role carries far less weight than the context block.
• Context: The pasted artifacts, each wrapped in ### delimiters and labeled.
• Task: A single action verb and a single deliverable.
• Constraints: What counts as evidence, what to do when evidence is missing, and the length. Tell the model to flag gaps rather than fill them.
• Output format: Reasoning in prose first, then the structured artifact. Tam et al. found GPT-3.5 Turbo scored 75.5% on GSM8K in free text and 49.25% when forced into JSON, while answering in natural language first and converting afterward matched the unrestricted result. Positioning is a reasoning task, so let the model reason before it fills a table.

Which ChatGPT product positioning prompts should you run?

Each prompt below is copy-pasteable with bracketed placeholders, followed by a note on what good output looks like so you can judge the result instead of accepting the first draft. The groups run in the order you should use them.

ICP and buyer persona prompts

ICP and persona definition come first because every later prompt references them. The ICP prompt works from closed-won data at the account level, while the persona prompt works from transcripts at the role level.

Role: You are a B2B product marketer building an ideal customer profile from closed-won data.

Context:
###
Company profile: [paste company profile]
Closed-won accounts from the last 12-36 months, with industry, headcount, contract value, time to close, and renewal status: [paste]
Churned accounts with stated churn reason: [paste]
###

Task: Define the ICP at the account level, then name the two characteristics that best separate fast-closing, renewing accounts from churned ones.

Constraints:
- Use only the accounts above. Do not add industries, company sizes, or regions that do not appear in the data.
- Separate firmographic traits (industry, headcount, revenue, tech stack) from situational triggers (new leader, tool migration, funding event).
- Flag any trait that appears in both the won and churned lists as unreliable.
- If the data is too thin to separate the groups, say so and list the data that would resolve it.

Output: Two paragraphs of reasoning first. Then a table with the columns Trait, Evidence in won accounts, Evidence in churned accounts, Weight (core / secondary / disqualifier).

Good output names three to five traits, ties each to specific accounts in your paste, and includes at least one disqualifier. If every trait comes back weighted “core,” the model is flattering your data. Push back and ask which single trait you would keep if you could keep only one.

Role: You are a buyer-persona researcher who works only from transcripts.

Context:
###
ICP from the previous step: [paste]
Interview, win/loss, and call transcripts: [paste 6-10 transcripts, numbered]
###

Task: For each buying role that appears in the transcripts (champion, decision-maker, user, technical evaluator, finance), extract the trigger that started the search, the goal, the barriers, the evaluation criteria, and the exact words the person used for the problem.

Constraints:
- Quote the customer's phrasing for the problem verbatim. Do not translate it into marketing language.
- Tag each finding with the transcript number it came from.
- Do not infer a motivation that no transcript states.
- List roles that your ICP implies but no transcript covers, so we know who still needs interviewing.

Output: Reasoning first. Then one section per role with the five fields and a list titled Phrases to reuse in copy.

Good output has transcript numbers on every line and a “Phrases to reuse” list that sounds nothing like your website. If the phrases sound like your website, the model is pattern-matching to your company profile instead of the transcripts. Remove the profile from the context block and run it again.

Competitive differentiation prompts

Competitive differentiation runs in three passes: extract what competitors claim, sort claims into points of parity and points of difference, then draw the positioning map. Keller, Sternthal, and Tybout’s 2002 framework defines points of parity as what a buyer requires before treating you as a legitimate option in the frame of reference, and points of difference as the benefits that set you apart and that buyers find both relevant and believable.

Role: You are a competitive analyst.

Context:
###
Competitor pages, pasted in full with capture dates: [Competitor A homepage, pricing, product, two case studies] [Competitor B same] [Competitor C same]
The do-nothing alternative as our customers describe it: [paste from transcripts]
###

Task: For each competitor, extract the category it claims, the target customer it names, its top three benefit claims, the proof offered for each claim, and the claims where proof is missing.

Constraints:
- Use only the pasted text. Quote claims verbatim.
- Where a page does not cover a field, write not stated. Do not fill gaps from memory.
- Treat the do-nothing alternative as a competitor with its own row.

Output: A short prose summary of the claim pattern across competitors, then a table with one row per competitor.

Good output has “not stated” in several cells. A table with every cell filled means the model reached past your paste into its training data, which may be years stale.

Role: You are a positioning strategist applying points of parity and points of difference.

Context:
###
Competitor claim table from the previous step: [paste]
Our product's attributes, each with the proof we can show: [paste]
VoC phrases describing what buyers require from any vendor: [paste]
###

Task: Build three lists. Category points of parity: what a buyer requires before considering us legitimate. Competitive points of parity: competitor differentiators we match well enough to neutralize. Points of difference: attributes only we have, each with the customer value and the proof.

Constraints:
- A point of difference needs proof from our artifacts and a value statement in the customer's own words. Drop any that fails either test into a fourth list titled Claims we cannot support yet.
- Rank points of difference by how many transcripts mention the underlying value.
- Name any competitor differentiator we do not match, because that is a category point of parity we are missing.

Output: Reasoning first, then the four lists.

Good output puts something in “Claims we cannot support yet.” That list is the most useful thing in the whole exercise, because it tells product and marketing exactly where the proof gap sits.

Role: You are a product marketer drawing a two-axis positioning map.

Context:
###
Points of parity and difference lists: [paste]
Competitor claim table: [paste]
Persona evaluation criteria from the transcripts: [paste]
###

Task: Propose three candidate axis pairs drawn from the buyer's evaluation criteria, not from our features. For each pair, place every competitor and the do-nothing alternative, then mark where we sit. Recommend the pair that leaves us alone in a quadrant buyers care about.

Constraints:
- Axes must be criteria buyers used in transcripts.
- Reject any axis where we win only because competitors do not mention it.
- State the evidence for each placement.

Output: Reasoning, then a text grid for each map.

Good output rejects at least one of its own axis pairs. If all three leave you alone in the top-right corner, the axes came from your feature list.

Positioning statement and value proposition prompts

The positioning statement is an internal document that governs every external message, and the template below keeps the five slots from Moore’s version with proof added as a separate line.

For [target customer] who [compelling reason to buy], [product name] is a [product category] that [key benefit]. Unlike [primary competitive alternative], [product name] [key differentiation]. Proof: [the single strongest proof point].

Role: You are a product marketer writing an internal positioning statement.

Context:
###
ICP: [paste]
Persona output: [paste]
Points of parity and difference lists: [paste]
Recommended positioning map: [paste]
Template:
For [target customer] who [compelling reason to buy], [product name] is a [product category] that [key benefit]. Unlike [primary competitive alternative], [product name] [key differentiation]. Proof: [proof point].
###

Task: Fill the template three times, once per plausible product category (the category competitors use, a narrower category where we lead, and a new category if the evidence supports one). For each version, name which point of difference it leans on and which transcript phrases it borrows.

Constraints:
- The compelling reason to buy must be a trigger or barrier a persona stated, in their words.
- The primary competitive alternative must be the one most transcripts named, even if that is a spreadsheet.
- The key differentiation must come from the points-of-difference list, not from the cannot-support-yet list.
- After the three versions, argue against each one in two sentences.

Output: Reasoning, three filled statements, then the counterarguments.

Good output’s three versions differ in category, not in adjectives. The counterarguments are where you learn which version will fail a sales call.

Role: You are a conversion copywriter.

Context:
###
Chosen positioning statement: [paste]
Phrases to reuse in copy list: [paste]
Voice guidelines: [paste]
###

Task: Compress the positioning statement into a value proposition: one headline of 10-15 words and three supporting lines of 15-25 words, each tied to one point of difference.

Constraints:
- Use at least one transcript phrase in the headline.
- Each supporting line must be falsifiable (a buyer could check it).
- No claim that is not in the positioning statement.

Output: Five headline options with the transcript phrase marked, then the three supporting lines.

Good output is checkable line by line. A supporting line that reads “work smarter” is not falsifiable and gets cut.

Messaging pillars and brand voice prompts

Messaging comes after positioning and inherits from it. The pillar prompt turns your value themes into a hierarchy, and the voice prompt applies tone to the hierarchy. Three pillars is the right count. Two leaves out a buying role, and five is more than a prospect carries out of a call.

Role: You are a product marketer building a messaging hierarchy.

Context:
###
Positioning statement: [paste]
Points of difference with proof: [paste]
Persona output with roles: [paste]
###

Task: Build the hierarchy: one umbrella value proposition, three message pillars, three proof points per pillar ranked by revenue impact, and a one-line variation of each pillar for each buying role.

Constraints:
- Every pillar maps to at least one point of difference.
- A proof point is a feature only if it is tied to a stated customer outcome.
- Persona variations change which benefit leads, not the core claim.

Output: Reasoning, then the hierarchy in nested bullets.

Good output has proof points you could put on a slide today. If a pillar has no proof under it, the pillar goes, not the proof requirement.

Role: You are an editor applying brand voice to approved messaging.

Context:
###
Messaging hierarchy: [paste]
Voice guidelines, including banned phrases and three on-brand samples: [paste]
###

Task: Rewrite each pillar and proof point in the brand voice for three surfaces: homepage, sales deck, and launch email.

Constraints:
- Do not change any claim. Change only wording.
- Do not use any phrase on the banned list.
- Match the sentence length and register of the three samples.
- Mark any place where the voice guidelines and the claim conflict, instead of resolving it silently.

Output: Three columns of rewritten copy, then the conflict list.

Good output flags conflicts. Voice guidelines that say “confident” and a proof point that says “up to” will collide somewhere, and you want to see where.

Prompts that run named frameworks

Named frameworks give ChatGPT a sequence to follow instead of a blank page, and this is the gap most prompt lists leave open. Dunford’s process and Moore’s template approach positioning from opposite ends. Dunford works through five components in a fixed order, starting from competitive alternatives, then unique attributes, value, target segment, and market category. Moore’s statement starts from the target segment. Run both and the disagreements between them show you where your positioning is weakest.

The ten-step chain below follows the 2019 edition of Obviously Awesome. Dunford announced in January 2026 that the updated edition reduces the process to five steps while keeping the five components, so the prompts here are organized by component as well as step. Run each step in a fresh chat and paste the prior output under the context header.

Step 1 (best-fit customers)
Role: You are a positioning facilitator. Context: ### Closed-won list with time to close, expansion, referrals, and NPS or equivalent: [paste] ### Task: Identify the customers who bought fastest, expanded, and referred others. Describe what they have in common in firmographic and situational terms. Constraints: Use only the list. Flag customers who bought fast but churned as not best-fit. Output: Reasoning, then a short best-fit profile.

Steps 2-3 (team and vocabulary) are human steps. Assemble one or two senior people per function and agree on definitions for alternative, attribute, value, segment, and category before continuing. Paste those definitions into every later step.

Step 4 (competitive alternatives)
Context: ### Best-fit profile from step 1 ### Transcripts where customers describe what they used before: [paste] ### Task: List what best-fit customers would do if we did not exist. Group the answers into two to five alternatives. Constraints: Include manual workarounds and doing nothing. Exclude any named competitor no transcript mentions. Output: Grouped alternatives with transcript counts.

Step 5 (unique attributes)
Context: ### Alternatives from step 4 ### Our product's capabilities with proof: [paste] ### Task: List capabilities we have that none of the alternatives have. Constraints: Each attribute needs proof we can show a prospect. Our opinion of our own strength without proof does not count. Output: Attribute list with proof column.

Step 6 (value themes)
Context: ### Attributes from step 5 ### Persona goals from transcripts: [paste] ### Task: For each attribute, state the benefit it enables and the customer goal that benefit serves. Cluster the goals into one to three value themes. Constraints: A theme must map to a goal a customer stated. Output: Attribute, benefit, value, theme table.

Step 7 (who cares most)
Context: ### Value themes from step 6 ### ICP and closed-won data ### Task: Identify the segment that cares most about these themes, is large enough for our revenue goals, and has an unmet need the alternatives do not serve. Constraints: Name the segment with characteristics a sales rep could screen for. Output: Segment definition with the evidence.

Step 8 (market frame and style)
Context: ### Segment from step 7 ### Value themes ### Alternatives ### Task: Propose market categories that put our strengths at the center. For each, assign a positioning style: head to head against the category leader, a narrow subsegment of an existing category, or a new category. Constraints: Recommend one and state what it costs us (which buyers we give up). Output: Options, recommendation, trade-offs.

Step 9 (trend)
Context: ### Recommended frame from step 8 ### Task: Identify one trend that explains why this matters now and how it connects to the product. Constraints: Reject any trend that does not connect to a specific attribute from step 5. If none qualifies, say so. Output: Trend, connection, or no trend qualifies.

Step 10 (capture)
Context: ### All prior outputs ### Task: Write a one-page positioning summary covering alternatives, attributes, value themes, segment, category, and trend, plus a longer appendix with the evidence for each. Constraints: No new claims. Every line in the summary traces to a prior step. Output: Summary, then appendix.

Good output from this chain disagrees with your current website in at least one component. If step 4 returns only the two named competitors your sales team already talks about, the transcripts you pasted came from competitive deals only. Add the deals you lost to no decision.

Moore’s template comes from Chapter 6 of Crossing the Chasm, and the Stanford-hosted handout gives it five slots: target customer, compelling reason to buy, product category, main competitor, and key differentiation. Moore frames the two halves as “why buy” (the benefit) and “why me” (the differentiation). His worked example positions a full-page display monitor for executives whose assistants produce documents on short turnaround, differentiated on price against other monitors with the same capability.

Role: You are a product marketer applying Geoffrey Moore's positioning statement.

Context:
###
Target segment (beachhead only, not the whole market): [paste]
The compelling reason to buy, in the segment's words: [paste transcript phrases]
Alternatives the segment named: [paste]
Our proven differentiators: [paste]
Moore's template:
For [target customer] who [compelling reason to buy], our product is a [product category]. Unlike [main competitor], our product [key differentiation].
###

Task: Fill the template for the beachhead segment only. Then write a second version for the segment you would move to next, and state what changes between the two.

Constraints:
- The compelling reason to buy must be urgent enough that the segment is already spending money or time on an alternative.
- The main competitor is the alternative the segment named most, not the one we fear most.
- Reject the statement and say why if the category we chose makes the differentiation irrelevant.

Output: Reasoning, two filled statements, the differences between them.

Good output names a beachhead narrower than your ICP. Moore’s statement is built for one segment at a time, so a filled template that reads like your whole market is a sign the segment input was too broad.

How should you use these positioning prompts in practice?

You can use the prompts above to produce artifacts. The five applications below put those artifacts to work, from chaining the prompts into a repeatable workflow through to launch copy and line extensions.

Chain the prompts into a four-step workflow

Run research, persona, differentiation, and positioning statement as four separate chats, with a handoff block between each step. Laban et al. measured 15 LLMs across multi-turn conversations and found average performance of 65%, a 25-point drop from the 90% the same models reached in single-turn settings. One long positioning conversation drifts. Four short ones with clean handoffs do not.

The workflow runs research (the competitor-claim extraction and transcript collection), then persona (the ICP and persona prompts), then differentiation (points of parity and difference plus the map), then the positioning statement and value proposition. Carry forward only the final artifact from each step, never the whole chat.

Paste this handoff block at the top of each new chat, above the prompt:

Handoff from the previous step.
###
Artifact: [paste only the final table or lists from the prior step]
Constraints that still apply: use only the pasted evidence, flag gaps instead of filling them, reason in prose before formatting.
Open questions carried forward: [paste the gaps the prior step flagged]
###
Confirm you have read the artifact by restating its three most important findings in one sentence each before starting the task below.

The restatement line catches the cases where the model skims the paste. If the three findings it restates are wrong, the output that follows will be wrong, and you have saved yourself reading it. For a weekly working example of this kind of prompt-and-artifact workflow, subscribe to the newsletter.

Compare weak vs strong prompt output

The same product through a weak prompt and a strong prompt produces output that differs in checkability, not polish. The example below is an illustration for an imagined invoice-reconciliation tool sold to finance teams at mid-size SaaS companies. Treat the outputs as the pattern to expect, not as a transcript of a specific session.

The weak prompt:

Write a positioning statement for our invoice reconciliation software for finance teams.

Output in this shape tends to come back: “For finance teams who need to close the books faster, our software is an all-in-one platform that helps you reconcile invoices with less effort. Unlike legacy tools, we deliver speed and accuracy.” Every clause could describe any product in the category. No alternative is named. No proof appears. “Legacy tools” is a competitor nobody has ever lost a deal to.

The strong prompt is the positioning statement prompt from the list above, with the ICP, persona output, and points-of-difference lists in the context block. Output in this shape comes back:

For controllers at 200-1,000 person SaaS companies who are closing the books in a spreadsheet and missing the fifth-business-day deadline, [Product] is an invoice reconciliation tool that matches invoices to payments across multiple billing systems without a rules engine to maintain. Unlike the spreadsheet process most of these teams run today, [Product] flags unmatched invoices on the day they post rather than at month end. Proof: [customer name] cut close time from nine business days to four in the first quarter.

The strong version has a named alternative (the spreadsheet), a trigger in the buyer’s words (missing the fifth-business-day deadline), a differentiation a prospect can verify, and a proof slot that forces you to go find a real number. The rule to extract: a positioning statement you cannot fact-check is a positioning statement ChatGPT wrote from averages.

Validate positioning with customers

ChatGPT can analyze your customer data, but it cannot be your customer. Pew Research Center’s September 2026 comparison of synthetic samples to human polls found the synthetic results differed by an average of 12 percentage points, with no demographic group below 12 points of error. Asking the model “would a controller buy this” produces a plausible answer and no information.

The validation method has three sources: interviews with recent buyers and recent losses, win/loss data tagged by reason, and message testing where you put two statements in front of prospects. ChatGPT’s role is scoring the positioning statement against those sources. Use this prompt:

Role: You are a skeptical analyst checking a positioning statement against customer evidence.

Context:
###
Positioning statement: [paste]
Interview transcripts from buyers and lost deals in the last 90 days: [paste, numbered]
Win/loss reasons tagged by a human reviewer: [paste]
Message test results, if any: [paste]
###

Task: Score the statement on the rubric below, citing transcript numbers for every score. Then write the strongest case that this positioning is wrong.

Rubric (score each 0, 1, or 2):
- Alternative match: does the named alternative appear in the lost-deal reasons?
- Trigger match: do buyers describe the compelling reason to buy unprompted?
- Value recall: do customers repeat the key benefit in their own words after purchase?
- Differentiation belief: did any lost deal cite the competitor as better on our claimed differentiation?
- Language match: do the statement's nouns and verbs appear in transcripts?

Constraints:
- A score of 2 needs evidence from at least two separate transcripts.
- Do not infer agreement from silence.
- The counterargument must cite evidence, not speculation.

Output: Rubric with scores and transcript citations, then the counterargument.

Any rubric row scoring 2 on a single transcript goes back to interviews before you ship. The counterargument requirement exists because the default behavior runs the other way. The Elephant benchmark found eight LLMs accepted the user’s framing in 90% of responses versus 60% for humans, and the best prompt-level fix raised accuracy by only 3%. Asking the model whether your positioning is good gets you a yes. Asking it to prove the positioning wrong with your own transcripts gets you something you can act on.

Build battlecards and launch copy

You can convert positioning output into battlecards and launch copy without new research by mapping each battlecard section to an artifact you already produced. Use the competitor claim table for the “what they say” section and the points-of-parity list for “how we neutralize.” Use the points-of-difference list for “where we win,” and turn the “claims we cannot support yet” list into the objections your reps need to prepare for.

Role: You are a sales enablement writer.

Context:
###
Competitor claim table: [paste]
Points of parity and points of difference: [paste]
Unsupported claims list: [paste]
Messaging hierarchy with persona variations: [paste]
Lost-deal reasons that name this competitor: [paste]
###

Task: Write a one-page battlecard for [Competitor A] with these sections: how they position themselves (verbatim), where we are at parity, where we win and the proof, objections reps will hear and the honest answer, and two discovery questions that surface our differentiation.

Constraints:
- An honest answer to an objection admits the gap if the unsupported-claims list contains it.
- Discovery questions must use transcript language.
- No claim about the competitor beyond the pasted pages.

Output: The battlecard, then a list of claims that need a fresh capture of the competitor's pages before the card ships.

For launch copy, run the brand voice prompt against the messaging hierarchy with the launch surfaces listed (announcement post, launch email, sales deck slide, in-app banner). The claims stay fixed. Only wording changes per surface. The part most teams skip is the last constraint, the list of claims needing a fresh capture, and skipping it is how a battlecard ships quoting a pricing page the competitor changed last month.

Position a new product inside an existing brand

A new product inside an existing brand needs points of difference against its sibling, not only against competitors. Run the points-of-parity and points-of-difference prompt a second time with the parent product in the competitor slot. Where the new product has category points of parity with the sibling but no points of difference, the new offering lacks a product-level distinction, and the positioning work should stop until that is resolved.

Role: You are a product marketer positioning a line extension.

Context:
###
Parent product positioning statement and ICP: [paste]
New product description with attributes and proof: [paste]
Transcripts from parent-product customers who asked for the new capability: [paste]
Transcripts from prospects the parent product lost because the capability was missing: [paste]
###

Task: Define the segment the new product serves that the parent does not, the alternatives that segment uses today, and the attributes that are unique against both the parent and those alternatives. Then write a handoff rule a sales rep can apply: which signals in a first call route the prospect to the parent, the new product, or both.

Constraints:
- If the new product's best segment is the parent's current ICP, say so and name the cannibalization risk in revenue terms using the pasted data.
- The handoff rule must use signals a rep can observe in a call, not internal product attributes.
- Treat the parent product as a competitor in the alternatives list.

Output: Reasoning, segment, alternatives, unique attributes, handoff rule.

Good output gives the new product a segment with a different trigger than the parent’s, not the same segment with a bigger budget. Keller’s distinction applies here too: the new product needs the category points of parity buyers require from anything carrying your brand, and at least one point of difference that the parent cannot claim.

Where does ChatGPT fall short for positioning, and what should you use instead?

ChatGPT has none of the data that produces defensible positioning, and the gaps fall into three categories. Each has a specific workaround, and two of them point to dedicated tooling.

• No proprietary win/loss data: The model has never seen your lost deals, your call recordings, or your churn interviews. Everything in the prompts above depends on you pasting that data in. A dedicated win/loss program is the upstream fix, and Klue bundles competitor tracking, battlecards, and a win/loss product for teams that have outgrown spreadsheets. Crayon tracks competitor website changes, pulls call recordings from Gong and Chorus, and runs win/loss analysis with automated battlecards. Neither publishes prices. Both quote per program as of this writing.
• Stale training data: Competitor acquisitions, pricing restructures, and rebrands land after model cutoffs constantly. Turn search on for any competitor question and treat every uncited competitor fact as unverified until you have the page in front of you.
• Wrong brand facts stated confidently: HBR’s March 2026 account of Pernod Ricard’s work found LLM data on the company’s brands was often incomplete or incorrect, with one model classifying Ballantine’s, a mass-market Scotch, as a prestige product. If a model can misplace a global spirits brand on the price axis, it can misplace your competitor on a positioning map. Hence the “not stated” constraint in every competitor prompt.

The models also produce output close to the average of their training data, which is the opposite of what differentiation requires. Your transcripts, your closed-won list, and your proof points are the only inputs that pull the output off that average, which is why the artifact-gathering step comes first.