Email the lab

An AI research lab that reads the logs.

Brainwork is the independent research lab run by Misha Manko. It operates an instrumented network of 50 real websites, records every visit from every AI crawler, measures which sources AI answers actually cite, and turns the data into published findings and working instruments.

Research network
50 sites, 9 industries
Logging
200+ days, continuous
Run by
One operator, no account managers
See the research programs Read the findings
access.logillustrative sample
  1. GPTBot/feed.xml200
  2. ClaudeBot/research/schema-markup200
  3. PerplexityBot/sitemap.xml200
  4. OAI-SearchBot/pricing200
  5. CCBot/200
  6. ChatGPT-User/feed.xml200
  7. Every AI bot/llms.txt0 requests

Across the network, every AI fetcher read the RSS feed before it read the homepage. Not one requested llms.txt.

Three questions, measured continuously.

Each program has its own instrument, its own data and its own publication trail. None of them rely on vendor dashboards or third-party estimates.

Running since 2025, from edge logs How AI crawlers read the web 50 real websites across 9 industries, every request logged at the edge. Which fetchers visit, what they open first, how often they return and what they ignore. This is the backbone of everything else the lab publishes.
Running since 2026, Common Crawl and the web graph What the open web index holds Common Crawl feeds most training corpora. The lab measures how much of a domain each crawl actually captured, when robots.txt started blocking it, how far the stored copy has drifted from the live page, and where the domain sits in the 118-million-node web graph.
Running since 2026, nightly, raw responses kept Which sources AI answers cite Nightly runs of real buyer prompts through ChatGPT and Google AI Overviews, with every cited source stored raw. The question is not "are we mentioned" but "which pages earn citations, and what do they have in common".

Tools the lab built because none existed.

Each one started as a measurement problem inside a research program. When the data needed a tool, the lab built it and kept it running.

Free web tool AI Visibility Snapshot Enter a website and see what ChatGPT, Claude, Perplexity and Google AI meet when they visit: bot access at the firewall, robots.txt, speed, text before JavaScript, feed, sitemap and structured data. Run the snapshot Free, no signup Common Crawl Checker Enter a domain and get seven panels: captures per crawl, robots.txt block history with dates, stored-copy drift, a live CCBot probe, sitemap coverage gaps and the domain's crawl-priority rank in the full web graph. Open the checker Free WordPress plugin AI Bot Access Checker Shows which AI bots can reach a WordPress site, and which robots.txt line, setting, security plugin or firewall rule turns the others away, with steps to fix it. Get the plugin
In service, client engagements only AI Visibility Tracker Runs a client's real buyer prompts through ChatGPT and Google AI Overviews every night, stores every raw response and reports which sources were cited. Built after testing the commercial trackers and finding the numbers mostly noise.
macOS app, on sale Quota A desktop widget that keeps AI usage limits in the corner of the screen. Built for the lab's own heavy use of coding agents, then productized for anyone who hits the same walls. quotawidget.com

Instrument, log, publish.

  1. Instrument

    Every claim starts with a sensor. Real sites, real edge logs, real API responses stored raw. If it cannot be measured, the lab does not have an opinion on it yet.

  2. Log

    Continuously, not as a one-off study. AI fetchers change behaviour month to month. A snapshot from last quarter is already a historical document.

  3. Publish

    Findings go out in the open with the method attached, including the ones that contradict the industry's favourite advice. The next person should not have to repeat the work.

House rules

  • Measurement before enthusiasm. AI is a tool, sometimes the right one, often not.
  • No guarantees about opaque systems. Anyone promising AI citation rankings is uninformed or lying.
  • Vendor dashboards are inputs to check, never sources of truth.
  • Client data never appears in published work. Findings are aggregated across the network.
  • Tools are built only when a research program needs one. Nothing speculative.
  • One operator, no account managers. You talk to the person who ran the query.

What the logs said.

Published research from the network, written up on Misha Manko's site. Method and data described in every piece.

N. 39 September 2026
N. 38 September 2026
N. 37 September 2026
N. 35 May 2026
N. 02 April 2026
N. 01

What AI bots actually read on your website

Report, the network's flagship study
April 2026
All research RSS feed

Bring the lab a hard question.

Research collaborations, data questions and consulting on the harder cases. One business day to reply. Productized audits and engagements run through mishamanko.com.

The AI visibility audit About Misha