KnitKnot
You are reading the agent-optimized layer of this page: the literal markdown we serve to AI crawlers and assistants, shipped in the page source of every visit. Making sure AI reads the right facts about a company is literally what KnitKnot does.

# KnitKnot Blog

Technical notes, product decisions, experiments, and company updates from the people building KnitKnot.

  • - Why AI recommends only 4-7 vendors in your category (2026-08-14)
    Google returns 10 results. AI usually returns 4-7 names. There's no page two. If you're not in the set, you're not in the evaluation.
    Markdown mirror: https://knitknot.ai/blog/ai-recommends-only-4-7-vendors.md
  • - Most of your citations do nothing (2026-08-03)
    The average AI answer cites 11.6 sources. Across 76,301 measured claims, fewer than 3 of them actually carry anything the buyer reads. We score every owned page on what its citations earn and trace every false or damaging claim back to the page that supplied it.
    Markdown mirror: https://knitknot.ai/blog/most-of-your-citations-do-nothing.md
  • - AI is citing pages that don't exist (2026-07-25)
    We probed every URL that ChatGPT, Claude, Perplexity, and Gemini cited across nearly 10,000 buyer-question answers. One citation in eighteen points at a page that is permanently gone, and nearly a third of answers lean on at least one.
    Markdown mirror: https://knitknot.ai/blog/ai-cites-pages-that-dont-exist.md
  • - The hidden cost of AI misinformation (2026-07-15)
    When AI gets a fact wrong about your company, it doesn't show up in your CRM as a lost deal. It shows up as a deal that never existed. We tried to quantify what that costs.
    Markdown mirror: https://knitknot.ai/blog/hidden-cost-of-ai-misinformation.md
  • - 72% of brands have factual errors in AI responses (2026-06-28)
    We analyzed 33,000 AI evaluations across ChatGPT, Claude, Perplexity, and Gemini for 47 B2B companies. 72% had at least one verifiably wrong factual claim. The errors cluster into five predictable, fixable patterns.
    Markdown mirror: https://knitknot.ai/blog/72-percent-brands-have-factual-errors.md
  • - The 10 questions AI buyers ask that your website can't answer (2026-06-18)
    We generate benchmark prompts grounded in real Google search data, with search volume attached to each one. The questions buyers ask ChatGPT, Claude, Perplexity, and Gemini are more adversarial, more specific, and more comparative than anything your website was designed to handle. Here are the ten patterns that show up most.
    Markdown mirror: https://knitknot.ai/blog/ten-questions-ai-buyers-ask.md
  • - Why AI recommends your competitor instead of you (2026-06-12)
    We analyzed 33,000 AI evaluations across four models. The most surprising finding: models disagree with each other on who to recommend 48.6% of the time. Which model the buyer opens matters more than most companies realize.
    Markdown mirror: https://knitknot.ai/blog/why-ai-recommends-your-competitor.md
  • - What ChatGPT says when a buyer asks to compare you (2026-06-10)
    We ran the same comparison prompt across ChatGPT, Claude, Perplexity, and Gemini for a B2B company. Four models gave four different answers. Two got the pricing wrong. One recommended the competitor based entirely on the competitor's own blog post.
    Markdown mirror: https://knitknot.ai/blog/what-chatgpt-says-when-buyers-compare.md
  • - Not all citations are equal (2026-06-06)
    A source that shaped the AI's recommendation carries more weight than one that provided a background fact. We model which sources had the most influence over what the buyer heard.
    Markdown mirror: https://knitknot.ai/blog/citations-are-ownership-claims.md
  • - AI is lying about your company (2026-06-03)
    We pulled every factual claim from our first 2,000 benchmark evaluations and checked them against reality. The error rate was higher than we expected, and the errors weren't random.
    Markdown mirror: https://knitknot.ai/blog/ai-is-lying-about-your-company.md
  • - A customer told us our benchmark was rigged (2026-05-28)
    We designed adversarial prompts to show companies where AI was misrepresenting them. Customers kept getting defensive about the prompts themselves. So we rebuilt the whole thing around real buyer behavior.
    Markdown mirror: https://knitknot.ai/blog/rebuilding-prompt-generation.md
  • - Prompt libraries are coverage optimization problems (2026-05-21)
    A bigger prompt library doesn't mean a better benchmark. We had hundreds of prompts and still missed the buyer situations that mattered most.
    Markdown mirror: https://knitknot.ai/blog/prompt-libraries-are-coverage-optimization-problems.md
  • - Approximating the Claude Engine (2026-05-14)
    ChatGPT, Perplexity, and Gemini all have incognito search. Claude doesn't. To benchmark how Claude represents companies, we had to find a way that respects Anthropic's terms instead of working around them. Here's what we built.
    Markdown mirror: https://knitknot.ai/blog/approximating-the-claude-engine.md
  • - Confident lies are worse than hedged ones (2026-05-07)
    Accuracy and conviction are independent axes. Most AI benchmarks only measure the first one. We model the interaction between what AI knows and how sure it sounds.
    Markdown mirror: https://knitknot.ai/blog/confident-lies-are-worse-than-hedged-ones.md
  • - What a candidate asks AI about your company (2026-04-29)
    A senior leader at a mid-size company asked us to track how AI describes them to candidates weighing offers. It wasn't our use case. With barely any changes, it worked.
    Markdown mirror: https://knitknot.ai/blog/brand-health-from-a-recruiting-question.md
  • - We stopped asking AI who wins (2026-04-22)
    Most LLM-as-judge systems ask one question: who's better? We decompose into structured signals and derive the outcome deterministically. Here's why.
    Markdown mirror: https://knitknot.ai/blog/we-stopped-asking-ai-who-wins.md
  • - Introducing KnitKnot (2026-04-14)
    KnitKnot runs the questions software buyers ask across four AI engines, saves every answer, checks the claims, and turns repeated problems into work a company can ship.
    Markdown mirror: https://knitknot.ai/blog/introducing-knitknot.md
  • - Why we pivoted KnitKnot (2026-04-07)
    We started KnitKnot as a digital sales room. Buyers liked it, but nobody needed it. Then we began asking what a sales room should look like when the buyer is an AI agent.
    Markdown mirror: https://knitknot.ai/blog/why-we-pivoted.md

Full text of everything above in one file: https://knitknot.ai/llms-full.txt

Raw mirror of this content: /llms.txt. Site-wide summary: /llms.txt · full content: /llms-full.txt

KnitKnot intelligence

Field notes.

What we are building, testing, measuring, and learning about the answers AI agents assemble.

A demand map connecting buyer questions, search signals, and comparison topics
Featured evidence plate

Aug 14, 2026 · 5 min read

Why AI recommends only 4-7 vendors in your category

Google returns 10 results. AI usually returns 4-7 names. There's no page two. If you're not in the set, you're not in the evaluation.

Read the field note

The archive

17 notes
  1. A source trail linking claim fragments to citations and evidence receiptsPlate 02sources

    Aug 03, 2026 · 16 min

    Most of your citations do nothing

    The average AI answer cites 11.6 sources. Across 76,301 measured claims, fewer than 3 of them actually carry anything the buyer reads. We score every owned page on what its citations earn and trace every false or damaging claim back to the page that supplied it.

  2. A source trail linking claim fragments to citations and evidence receiptsPlate 03sources

    Jul 25, 2026 · 9 min

    AI is citing pages that don't exist

    We probed every URL that ChatGPT, Claude, Perplexity, and Gemini cited across nearly 10,000 buyer-question answers. One citation in eighteen points at a page that is permanently gone, and nearly a third of answers lean on at least one.

  3. Layered answer fragments showing conflicting representations of the same companyPlate 04AI presence

    Jul 15, 2026 · 6 min

    The hidden cost of AI misinformation

    When AI gets a fact wrong about your company, it doesn't show up in your CRM as a lost deal. It shows up as a deal that never existed. We tried to quantify what that costs.

  4. Layered answer fragments showing conflicting representations of the same companyPlate 05research

    Jun 28, 2026 · 10 min

    72% of brands have factual errors in AI responses

    We analyzed 33,000 AI evaluations across ChatGPT, Claude, Perplexity, and Gemini for 47 B2B companies. 72% had at least one verifiably wrong factual claim. The errors cluster into five predictable, fixable patterns.

  5. A demand map connecting buyer questions, search signals, and comparison topicsPlate 06AI presence

    Jun 18, 2026 · 6 min

    The 10 questions AI buyers ask that your website can't answer

    We generate benchmark prompts grounded in real Google search data, with search volume attached to each one. The questions buyers ask ChatGPT, Claude, Perplexity, and Gemini are more adversarial, more specific, and more comparative than anything your website was designed to handle. Here are the ten patterns that show up most.

  6. Layered answer fragments showing conflicting representations of the same companyPlate 07research

    Jun 12, 2026 · 9 min

    Why AI recommends your competitor instead of you

    We analyzed 33,000 AI evaluations across four models. The most surprising finding: models disagree with each other on who to recommend 48.6% of the time. Which model the buyer opens matters more than most companies realize.

  7. Layered answer fragments showing conflicting representations of the same companyPlate 08benchmarks

    Jun 10, 2026 · 8 min

    What ChatGPT says when a buyer asks to compare you

    We ran the same comparison prompt across ChatGPT, Claude, Perplexity, and Gemini for a B2B company. Four models gave four different answers. Two got the pricing wrong. One recommended the competitor based entirely on the competitor's own blog post.

  8. A source trail linking claim fragments to citations and evidence receiptsPlate 09sources

    Jun 06, 2026 · 6 min

    Not all citations are equal

    A source that shaped the AI's recommendation carries more weight than one that provided a background fact. We model which sources had the most influence over what the buyer heard.

  9. Layered answer fragments showing conflicting representations of the same companyPlate 10AI presence

    Jun 03, 2026 · 8 min

    AI is lying about your company

    We pulled every factual claim from our first 2,000 benchmark evaluations and checked them against reality. The error rate was higher than we expected, and the errors weren't random.

  10. A governed page revision bounded by approved facts, review marks, and release recordsPlate 11methodology

    May 28, 2026 · 7 min

    A customer told us our benchmark was rigged

    We designed adversarial prompts to show companies where AI was misrepresenting them. Customers kept getting defensive about the prompts themselves. So we rebuilt the whole thing around real buyer behavior.

  11. A demand map connecting buyer questions, search signals, and comparison topicsPlate 12benchmarks

    May 21, 2026 · 5 min

    Prompt libraries are coverage optimization problems

    A bigger prompt library doesn't mean a better benchmark. We had hundreds of prompts and still missed the buyer situations that mattered most.

  12. A frozen baseline and later observation arranged as a measured evaluation windowPlate 13methodology

    May 14, 2026 · 5 min

    Approximating the Claude Engine

    ChatGPT, Perplexity, and Gemini all have incognito search. Claude doesn't. To benchmark how Claude represents companies, we had to find a way that respects Anthropic's terms instead of working around them. Here's what we built.

  13. A frozen baseline and later observation arranged as a measured evaluation windowPlate 14measurement

    May 07, 2026 · 6 min

    Confident lies are worse than hedged ones

    Accuracy and conviction are independent axes. Most AI benchmarks only measure the first one. We model the interaction between what AI knows and how sure it sounds.

  14. A demand map connecting buyer questions, search signals, and comparison topicsPlate 15product

    Apr 29, 2026 · 4 min

    What a candidate asks AI about your company

    A senior leader at a mid-size company asked us to track how AI describes them to candidates weighing offers. It wasn't our use case. With barely any changes, it worked.

  15. A frozen baseline and later observation arranged as a measured evaluation windowPlate 16scoring

    Apr 22, 2026 · 8 min

    We stopped asking AI who wins

    Most LLM-as-judge systems ask one question: who's better? We decompose into structured signals and derive the outcome deterministically. Here's why.

  16. A governed page revision bounded by approved facts, review marks, and release recordsPlate 17product

    Apr 14, 2026 · 6 min

    Introducing KnitKnot

    KnitKnot runs the questions software buyers ask across four AI engines, saves every answer, checks the claims, and turns repeated problems into work a company can ship.

  17. A governed page revision bounded by approved facts, review marks, and release recordsPlate 18launch

    Apr 07, 2026 · 7 min

    Why we pivoted KnitKnot

    We started KnitKnot as a digital sales room. Buyers liked it, but nobody needed it. Then we began asking what a sales room should look like when the buyer is an AI agent.