TL;DR
Feed an AI your own raw research and force a citation tag after every cell, instead of asking it to “compare” companies from memory. Use a deep-research agent (ChatGPT Deep Research, Claude Research, or Gemini 3.1 Pro Deep Research) only to collect and structure the evidence; you still verify pricing and feature claims against primary sources. The prompt below produces a markdown matrix where every cell reads either [note_id] or [not in notes], so a stakeholder asking “where did this number come from?” gets a real answer.
The task
You are preparing a pitch deck, product strategy memo, sales battlecard, or website comparison page. You need a side-by-side matrix of your product versus 3-8 competitors across pricing, features, positioning, and customer fit. The hard part is not building the grid. It is making sure that every cell can be traced to a source, because the first thing a skeptical exec does is point at one number and ask where it came from.
The trap with general AI chat is that it will happily fill the whole grid from training data that is months stale and was never accurate about a private competitor’s pricing in the first place. Web search reduces that risk but does not remove it: a model can still misread a pricing page, misattribute a feature to the wrong tier, or extrapolate past what the source actually says. Citations are your verification mechanism, not a formatting nicety.
When AI is the right tool
- You already have raw research (pricing pages, G2 reviews, sales-call transcripts) and need help structuring it into one grid.
- You want a first-pass matrix that a human then verifies cell by cell.
- You are tracking a fast-moving category and need a refresh every 4-8 weeks.
- You need the same comparison in two framings (sales-facing battlecard vs. board-facing strategy memo).
When not to rely on AI alone
- Exact pricing and feature checkmarks. These must come from primary sources you can link.
- Acquisition rumors, layoffs, or unreleased roadmap. Models will state these as fact with full confidence.
- Regulated industries (healthcare, fintech) where a wrong claim in a published comparison creates legal exposure.
Pick the right research tool
As of June 2026, all three major assistants ship a deep-research agent that plans searches, follows links, and returns a cited report. They differ in run length, citation discipline, and what they can export. For a comparison matrix the export format matters, because you want to land in a spreadsheet, not a wall of prose.
| Tool (June 2026) | Deep-research run | Citation behavior | Best for the matrix step |
|---|---|---|---|
| ChatGPT Deep Research (GPT-5.5) | Up to ~30 min, longest reports; 25-250 runs/month by plan | Inline source list; leads BrowseComp agentic-browsing | Broadest source sweep, longest first draft |
| Claude Research (Opus 4.7) | Fewer pages, careful synthesis | Citations are contractually required on Claude’s API web-search tool; strong on ambiguous sources | Highest trust per cell, least over-extrapolation |
| Gemini 3.1 Pro Deep Research | BrowseComp ~85.9; exports to Google Sheets / Excel / Docs | Source list plus Canvas visuals | One-tap export to a live spreadsheet matrix |
A practical pattern: run the sweep in ChatGPT or Gemini to gather sources fast, then paste the raw notes into the citation-locked prompt below (Claude Opus 4.7 is the safest model for the “do not infer” step). Whatever tool you use, the discipline is the same: the model structures evidence you supply, it does not invent it.
What to feed the AI
- The competitor list with their pricing-page URLs.
- 5-8 dimensions to compare (pricing tiers, integrations, target customer, support SLA, deployment model, etc.).
- Raw notes per competitor: copy-pasted pricing pages, three G2 or Capterra reviews each, one analyst quote.
- Your product’s honest positioning, so the framing is fair rather than flattering.
- The required citation format: a bracketed tag after every cell.
Copy-ready prompt
You are a competitive intelligence analyst.
Build a comparison matrix of [our_product] vs. [competitor_list].
Dimensions to compare:
[dimension_list]
Raw notes per competitor are below. Use ONLY these notes. Do not infer.
"""
[raw_notes]
"""
Output:
1. A markdown table. Rows = dimensions, columns = each company.
2. After each cell, add a citation tag in brackets: [note_id] or "[not in notes]".
3. Below the table, list:
- 3 insights (patterns across competitors)
- 3 white-space gaps (dimensions where no one is winning)
- 3 risks for [our_product] based on the data
Rules:
- If a cell is not supported by the notes, write "not in notes".
- Do not estimate pricing. Do not infer headcount.
- Do not use marketing adjectives.Use [bracket] placeholders rather than curly braces, and swap in your real values before sending. The """ fences keep your raw notes from being read as instructions.
Recommended output structure
A clean markdown matrix with a citation tag in every cell, followed by an insights block (patterns, gaps, risks) of 200-300 words. Keep visual decoration minimal. This is a working document, not a pitch slide. If you used Gemini, export the table straight to Sheets so the refresh date and per-cell source live in columns you can sort.
How to check the output
- Click into 30% of cells at random and verify each against its source URL.
- Flag every
[not in notes]cell and decide whether it needs primary research. - Stress-test each insight: would your most skeptical colleague accept it from this data alone?
- If you can reach a competitor’s actual customer, have them sanity-check the row about their tool.
Common mistakes
- Letting the model fill pricing or feature gaps with plausible guesses instead of writing
[not in notes]. - Comparing on dimensions that do not matter to your actual buyer.
- Building a matrix that flatters your product instead of exposing real gaps.
- No refresh cadence, so last quarter’s matrix is already wrong on at least one price.
Keep it a living document
Turn the matrix into a living document with a refresh date in the header. After every won or lost deal, log which row actually decided it. Use that record to prune low-signal dimensions and add the ones buyers keep asking about. A 6-row matrix the buyer cares about beats a 20-row one nobody reads.
FAQ
- How many competitors is too many? 3-5 for a sales battlecard, 6-10 for a strategy memo. Past 10, the signal dilutes and nobody scans the whole grid.
- Should I include indirect competitors? Yes, in a separate section. Buyers compare against alternatives (spreadsheets, in-house builds, doing nothing) that you may not treat as direct rivals.
- Which model should I use for the citation-locked step? As of June 2026, Claude Opus 4.7 is the safest for “use only these notes,” since citation is enforced at the API level and it over-extrapolates less. Use ChatGPT or Gemini for the wider source sweep first.
- How do I avoid bias toward my own product? Have someone outside marketing fill in your column first, before they see how competitors are described.
- Can I publish the matrix? Only after legal review. Even citing public sources can attract complaints, and a misattributed feature claim is the kind of thing competitors escalate.
Related
Pair this with competitor analysis AI workflows, the competitor feature matrix AI build, and a competitor content teardown for the marketing-message angle.