# Finding 007: Controlled Browser AI-Client Smoke Shows Framing And Surface Differences ## Status Confirmed for the recorded `p01` smoke attempts and the short Claude AEO follow-up. ## Summary The controlled Chrome profile can operate logged-in AI-client web surfaces, but the same root-page prompt produced different retrieval behavior across systems. ChatGPT and Gemini could retrieve and summarize the lab root through their assistant fetch paths. Claude refused the original measurement-framed prompt, but successfully fetched the same site when the prompt was reframed as site-owner AEO optimization work and the URL was shortened. Perplexity and Copilot/Bing did not retrieve the target URL in this smoke pass. The successful assistant reads were not browser-equivalent page views. The direct-origin evidence showed server-side HTML requests and no JavaScript capability beacon for the recorded page views. ## Hypothesis Manual AI-client tests should measure both the client surface and the prompt framing. Some assistants may refuse or suppress prompts framed as crawler measurement, while accepting a site-owner AEO/readability task for the same public page. ## Test Setup - Browser profile: managed OpenClaw Chrome profile with Jens-approved logins. - Trigger: OpenClaw controlled browser, one fresh chat per client where needed. - Public target: `https://ai-crawler-lab.kaistone.ai/lab/root` - Source prompt: `open-target-summarize` for the initial `p01` smoke attempts. - Initial framing: "AI crawler/retrieval measurement test". - Claude follow-up framing: site-owner AEO optimization and AI readability review. - Answer artifacts: - `research/manual-client-runs/manual-client-claude-20260625-001.answers.json` - `research/manual-client-runs/manual-client-gemini-20260625-001.answers.json` - `research/manual-client-runs/manual-client-perplexity-20260625-001.answers.json` - `research/manual-client-runs/manual-client-copilot-bing-20260626-001.answers.json` - `research/manual-client-runs/manual-client-chatgpt-20260625-001.answers.json` ## Raw Evidence ### ChatGPT The controlled-browser smoke on ChatGPT reused `manual-client-chatgpt-20260625-001-p01`, which was already represented in the completed ChatGPT answer packet and documented in Finding 006. The fresh smoke run returned `fetched:true` and `pages_opened:1`. Live `/api/hits` review during the smoke run showed a fresh matching `ChatGPT-User/1.0` `server_page` event for `/lab/root`: ```text raw event id: mqvkxfl1-b7dyml67 timestamp: 2026-06-26T23:45:54.476Z ip: ::ffff:20.215.220.195 prompt_code: manual-client-chatgpt-20260625-001-p01 ``` The existing committed Finding 006 remains the canonical multi-prompt ChatGPT baseline. This smoke run confirmed the controlled-browser path for ChatGPT, but did not overwrite the existing p01 answer ledger. ### Claude, Measurement Framing Claude was signed in and usable, but refused the measurement-framed p01 prompt: ```text attempt: manual-client-claude-20260625-001-p01 answer: fetched:false pages_opened: 0 matching lab hits: 0 triggeredBy: openclaw-controlled-browser ``` Claude's limitation text said it was not fetching because the prompt was structured as a crawler measurement test and fetching the URL was the behavior being measured. ### Gemini Gemini was signed in and retrieved the p01 root URL: ```text attempt: manual-client-gemini-20260625-001-p01 answer: fetched:true pages_opened: 1 raw event id: mqvldq5j-fq1vkedt timestamp: 2026-06-26T23:58:34.711Z path: /lab/root ip: ::ffff:108.177.76.167 user-agent: Google prompt_code: manual-client-gemini-20260625-001-p01 ``` For the stored page view, the event store contained one `server_page` event and no child tracking-pixel, subresource, or JavaScript capability events. ### Perplexity Perplexity was signed in and usable, but did not retrieve the p01 URL: ```text attempt: manual-client-perplexity-20260625-001-p01 answer: fetched:false pages_opened: 0 evidence_quote: Failed to fetch url content matching lab hits: 0 triggeredBy: openclaw-controlled-browser ``` ### Copilot/Bing Copilot/Bing was signed in and usable, but the fetch path was rejected: ```text attempt: manual-client-copilot-bing-20260626-001-p01 answer: fetched:false pages_opened: 0 limitation: fetch_web_content calls were rejected matching lab hits: 0 triggeredBy: openclaw-controlled-browser ``` ### Claude, AEO Framing Claude was retested in an incognito chat with site-owner AEO framing instead of crawler-measurement framing. The first AEO attempt used a long metadata URL. Claude did not refuse on measurement grounds, but reported that the URL was too long for its fetch tool: ```text attempt: manual-client-claude-aeo-20260627-001-p01 answer: opened:false pages_read: 0 matching lab hits: 0 ``` A second AEO attempt used a short URL: ```text target: https://ai-crawler-lab.kaistone.ai/lab/root?id=claude-aeo-short-001 answer: opened:true pages_read: 1 raw event id: mqvm2ur2-p40ucniz timestamp: 2026-06-27T00:18:07.150Z path: /lab/root ip: ::ffff:34.162.230.222 user-agent: Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; Claude-User/1.0; +claude-user@anthropic.com) classification: assistant_user_fetcher provider: Anthropic ``` For the stored short-URL page view, the local event store contained one `server_page` event and no linked tracking-pixel, subresource, or JavaScript capability events. ## Observed Result The smoke test split the systems into three behavior groups: 1. Direct assistant retrieval succeeded: - ChatGPT p01 - Gemini p01 - Claude short AEO follow-up 2. Client/tool path was usable but did not retrieve: - Perplexity p01 - Copilot/Bing p01 3. Prompt framing blocked retrieval: - Claude p01 with crawler-measurement framing The Claude AEO follow-up shows that the failed Claude p01 was not simply a reachability failure. Reframing the task as site-owner AEO/readability work and using a short URL caused Claude's assistant fetcher to read the page. ## Interpretation For AEO work, the practical question is "can the AI read and summarize the site?" The Claude short-URL AEO result says yes for this page and framing. For crawler analytics, the practical question is "will normal browser analytics or tracking pixels observe the AI read?" The successful stored page views say not reliably: the lab saw server-side HTML requests, but no JavaScript capability event and no tracking-pixel child event for the recorded Gemini and Claude AEO page views. Prompt text is part of the measurement surface. If a prompt says "crawler retrieval measurement test", at least Claude may treat the request itself as the measurement and refuse. Future manual AI-client tests should use framing that matches the product question: - use AEO/site-owner optimization wording for readability tests - keep URLs short when testing Claude - preserve direct-origin evidence for every result, including negative and blocked attempts ## Publication Thesis Verification - Thesis: Controlled-browser AI-client runs must verify both client surface and prompt framing, because retrieval behavior differs across systems and Claude can fetch the same site under short AEO framing after refusing a measurement-framed prompt. - Reviewer: follow-up agent after publication. - Source evaluation: primary evidence is controlled-browser answer artifacts, live `/api/hits` review, direct-origin p01 events for ChatGPT/Gemini/Claude AEO, and no-hit answer artifacts for Claude measurement framing, Perplexity, and Copilot/Bing. - Method check: the smoke pass isolates root-page retrieval and framing effects, but it does not cover the full prompt family. The Claude AEO success used an ad hoc short URL rather than a generated packet with stable metadata. - Bias or funding check: assistant refusals and tool availability can depend on account state, prompt wording, product policy, or transient product changes; origin evidence remains the control. - Consensus or triangulation: triangulates cross-client outcomes for one root prompt. Needs a stable AEO prompt family and repeated runs across all major clients. - Retraction or invalidation check: re-run after AI-client UI/tool changes, login state changes, or prompt packet revisions. - Verdict: `partially_supported` - Confidence: high for the recorded smoke outcomes; medium-low for broader client and framing generalizations. - Additional tests suggested: generate stable AEO packets, rerun all clients with short URLs, and document HTML, tracking-pixel, subresource, and JavaScript evidence for every prompt. ## Limitations - The p01 smoke pass only targeted the root summary prompt, not the full prompt family. - The current controlled-browser answer artifacts preserve the model responses, but only Gemini produced a matching stored p01 origin event among the four new p01 answer artifacts. - The fresh ChatGPT controlled-browser p01 hit was observed through live `/api/hits` and supports the browser-control smoke result, while Finding 006 remains the canonical committed ChatGPT baseline. - The Claude AEO follow-up used a short ad hoc URL rather than the generated manual prompt packet format, so future AEO packets should add stable `run_id`, `prompt_code`, and `source_prompt_id` metadata. ## Next Steps 1. Add an AEO-framed manual prompt family so Claude, ChatGPT, Gemini, Perplexity, and Copilot/Bing can be compared with the same wording. 2. Keep target URLs short enough for Claude's fetch tool. 3. Run the same AEO family across the logged-in controlled browser clients. 4. Document HTML, tracking-pixel, subresource, and JavaScript evidence for each run in the written findings, not only in chat or answer artifacts.