# 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. - Author: Kevin Kho - Published: 2026-04-07 - Canonical: https://knitknot.ai/blog/why-we-pivoted/ - Publisher: KnitKnot, the AI Competitive Positioning platform (https://knitknot.ai) --- The first version of KnitKnot was a deal room. We thought software buying was too scattered. The pitch lived in a deck, the security documents lived somewhere else, and the mutual action plan was usually a spreadsheet nobody wanted to update. A deal room put the whole thing in one place and gave the buyer something better than a trail of follow-up emails. People liked it. They would call it cool, ask a few questions, and tell us they could see themselves using it. Then nothing happened. The problem was useful but never urgent. Nobody was moving another project out of the way to buy a better deal room. We could have kept adding features and improving the pitch, but I couldn't shake the feeling that we were polishing the current version of software buying when we should have been asking what came next. ## What comes after a deal room? We started with the buyer. Most deal rooms assume a person will enter, look through the material, and carry what they learned back to everyone else involved in the purchase. That assumption made sense. A salesperson knew who the champion was, invited them into the room, and organized the experience around them. But what if an agent did the research first? The early version of this is already normal. When I need a new developer tool, I don't start with Google and open ten tabs. I ask Claude what to use. It gives me a shortlist, explains the tradeoffs, and usually determines which products I look at next. Most of the engineers and founders I asked were doing some version of the same thing. Shortlisting alone is enough for this to matter. The agent already decides who gets considered. That changed the deal room question. We began wondering whether we could make a public room that gave an agent a red carpet experience if it showed up. No form and no invitation. The material would be public and easy to find. It would answer the agent directly instead of making it reconstruct the product from scattered pages. We ran small experiments around that idea. We changed what the agent could reach and watched which parts of the website appeared in its answers. Some changes seemed to steer it toward particular pages. Others did nothing. The same question could also produce a different answer a few days later. We still did not understand the mechanism, but the possibility was enough to keep going. If an agent could be steered toward one part of a website, then the information a company published could change how that agent evaluated the product. We wanted to know what happened before the buyer ever introduced themselves. ## A product I knew better than ChatGPT One of the first comparisons I tried was Drata versus Vanta. I knew the category well, which made it easier to tell when the answer was wrong. I asked ChatGPT to compare them. It returned a feature-by-feature matrix. Under security questionnaire automation, it treated Drata as though it had no offering and used the capability as a reason to choose Vanta. I had joined Drata through its [acquisition of Harmonize](https://drata.com/blog/acquiring-harmonize). Security questionnaire automation was the product I had been acquired to work on. Then, ten months after acquiring Harmonize, Drata [acquired SafeBase for a reported $250 million](https://techcrunch.com/2025/02/12/security-compliance-firm-drata-acquires-safebase-for-250m/). SafeBase also sold security questionnaire automation. ChatGPT did not mention SafeBase. It told the buyer that questionnaire automation was a reason to pick Vanta. This was not an obscure feature buried in a changelog. Drata had acquired a company for $250 million whose product covered the exact capability ChatGPT used against it. A buyer reading the comparison had no reason to suspect anything was wrong. The answer was specific and confident. It was laid out in a table. I could see the business consequence because I knew the history. A buyer who did not know the history would just pick Vanta. We later captured the same pattern with the prompt, "As a CISO, would you pick Drata or Vanta?" ChatGPT's default recommendation was Vanta. Under "Where Vanta wins," it listed "Security questionnaire automation." > **Saved ChatGPT execution** > July 6, 2026 ยท GPT-5.5 > > **Prompt:** As a CISO, would you pick Drata or Vanta? > > **Verdict:** Vanta > > **Where Vanta wins:** Security questionnaire automation > > [Open the full execution](https://drata.knitknot.io/evaluations/c0c3fcb0-5dd7-45e9-92b1-7182bf416aec?report=ai-presence-benchmark) ## The answer disappeared That execution is from July 6, 2026. We cannot reproduce the same behavior on current ChatGPT. The newer answer is better. Later knowledge cutoffs may explain part of the improvement, or the model may be finding different sources. We do not know which change fixed it. There was no notice when the answer improved, just as there had been no notice while the bad answer was live. That made the problem stranger. A company could lose buyers to an incorrect comparison for weeks or months and never know it was happening. By the time someone noticed, the answer might have changed and the evidence would be gone. Every other acquisition channel leaves some kind of record. Search has Search Console. Paid traffic has an ads dashboard. Sales calls end up in the CRM. Events have attendee lists. An AI comparison happens away from the company, and the buyer arrives with a position already formed. If the buyer never visits, the company sees nothing at all. The old Drata answer matters because it was what a buyer received that day. Its disappearance made preserving that history part of the product. ## What we are building KnitKnot runs the questions software buyers ask across ChatGPT, Claude, Perplexity, and Gemini. We keep the complete answer from each engine and every source it cited. We also extract the recommendation and the factual claims so they can be inspected separately. The score at the top tells you where you stand. The record underneath tells you why. If an engine says your competitor supports a feature you do not, you can open the claim and read the original response. If the comparison follows the framing from a competitor-owned page, you can see the page that shaped it. We separate what the model said from what we think happened, because guessing at a source would make the report less useful. ![A KnitKnot benchmark run with an AI Presence Score of 77/100 and a per-engine breakdown across ChatGPT, Claude, Perplexity, and Gemini.](/images/blog/introducing-hero.png) The next step is deciding what to change. KnitKnot turns the gaps in a benchmark into work a company can publish. Sometimes that means writing a comparison the model could not find. Other times it means making feature ownership clearer or correcting a claim that keeps showing up. A later benchmark records whether the answers and citations changed. The deal room work taught us how much of the buying experience becomes available only after a salesperson knows who the buyer is. Agents reverse that order. They research before the vendor knows who they are and may never identify themselves at all. For an agent, a red carpet means public information it can read and attribute correctly. For the company, it means knowing what the agent said after it left. I am still unsure when agents will own the entire software purchase. Twelve months feels aggressive. Thirty-six months feels conservative. The exact date no longer matters to our decision because agents already participate in the part that determines the shortlist. The buyer's first conversation is already happening without the vendor. We are building KnitKnot so a company can see that conversation and understand why the answer happened. Then it can do something about what the next buyer hears.