Tailor research · November 2, 2026
What we learned testing Tailor with public AI agents
We tested public Claude, ChatGPT and Gemini sessions against Tailor's website. Claude needed fewer page reads in three comparisons when we supplied an HTML assessment with complete sources. We did not find reliable tool discovery across all agents.
Three comparisons with fewer page reads
Claude still answered the requested questions in these trials. We counted the page reads shown in its trace. We did not measure tokens, response time, unique server requests or cost savings.
| Task | Starting point | Ordinary pages | Evaluation |
|---|---|---|---|
| Shopify research | Current homepage | 6 | 3 |
| Webflow research | Current homepage | 10 | 1 |
| Shopify pricing holdout | Existing pricing page | 5 | 2 |
What the tests changed
- Short evidence excerpts omitted prerequisites and verification steps. The evaluator now includes the complete source pages and names the questions our team still needs to answer.
- A free Webflow Starter plan blocked the documented installation route. The assessment now flags it and withholds an experiment proposal.
- Pricing tables alone did not prevent arithmetic and plan errors. We added explicit tier explanations and visible inclusion labels; some agent mistakes remained.
- In a separate ten-trial rendering experiment, Gemini twice returned the initial price and missed JavaScript-added facts. Claude recovered the values by fetching script and JSON. Browser controls rendered the changes correctly.
Method and limits
The November 2 research round contained 44 fresh trials, including failed and incomplete runs, plus separately retained pilots and diagnostics. It compared a frozen pre-session site, current ordinary pages and guided evaluations, then used new requirements and pricing values for follow-up tests. The historical site already supported strong Claude and ChatGPT answers.
This is one site with small samples, unblinded scoring, possible retrieval caches and variable client behavior. A supplied URL does not test spontaneous discovery. HTTP tool use is distinct from native MCP or WebMCP use. Client health failures were retained separately. We did not measure conversion lift, completed purchases, booking completion or search-ranking improvement. The JavaScript fixtures were noindex and did not run the Tailor SDK.
We have not yet tested whether these changes increase citations across customer sites.
Download the aggregate experiment record (JSON)