AI Changed Who Visits Your Website. Your Cookie Banner Didn't Notice.

Cookie consent law was written for a person looking at a screen. In 2026, a meaningful share of your "visitors" are not people: they are training crawlers harvesting text for the next model, retrieval bots building AI search indexes, and — the newest arrival — full agent browsers operating a page on a human's behalf. Each breaks a different assumption your consent setup quietly relies on. This guide maps the new visitor landscape and what it changes for consent, analytics, your content, and your privacy documentation.

The new visitor taxonomy

• Training crawlers (GPTBot, ClaudeBot, Meta-ExternalAgent and peers) fetch content to train models. They generally respect robots.txt and crawl at scale.
• AI search / retrieval crawlers (OAI-SearchBot, Claude-SearchBot, PerplexityBot) build retrieval indexes so assistants can cite the live web.
• User-triggered fetchers (ChatGPT-User, Claude-User) fetch a specific page because a human asked a question about it. Vendors document that robots.txt may not govern these, on the argument that a human initiated the request.
• Agent browsers (ChatGPT Atlas, Perplexity Comet, Claude in Chrome and similar) run a *full browser*: they load your tags, render your banner, and click your buttons the way a person would.

Problem 1: agents click "Accept All", and that is not consent

• Your consent analytics are polluted. Agent accepts inflate your opt-in rate and your "consented" audience. If you steer marketing spend by consent-rate dashboards, you are partly measuring robots.
• Your consent records include non-consents. Under Article 7(1) you must be able to demonstrate consent; a record whose user-agent and behaviour pattern screams automation demonstrates the opposite.
• Defensive defaults help. Where automation is detectable (headless signals, navigator.webdriver, known agent user-agent strings), the conservative move is to treat the visit as consent-denied and serve the content without writing non-essential state. You lose nothing — an agent was never going to click your retargeting ad — and your records stay clean.

Problem 2: your content is training data, and robots.txt just became a legal document

• Consent is effectively off the table for scrapers. The EDPB's position is that a scraper with no relationship to the people whose data appears on your pages cannot realistically obtain valid consent — and, importantly, that *the absence of a robots.txt rule is not consent*.
• Legitimate interest is the battleground. Scraping AI companies must pass the three-part legitimate-interest test, and — this is the shift — the EDPB treats your machine-readable signals as weights in that balancing test: robots.txt directives, ai.txt, CAPTCHAs, login walls, and rate limits all now carry legal significance, not just technical effect.
• An anonymization test (no isolation, no linkage, no inference) decides whether trained models still hold personal data.

Problem 3: your own AI features drag your privacy docs along

• The chatbot must disclose it is AI (Art. 50(1)) — a label in the widget, not a footnote.
• AI-generated content needs disclosure and machine-readable marking, with a transitional deadline of 2 December 2026 for generative systems that were already on the market.
• Your privacy policy must cover the AI data flow. If chat transcripts go to a model provider, that provider is an Article 13 recipient and belongs on your subprocessor list, together with your training stance — does the provider train on your users' conversations? "We use AI" without naming where the data goes is the new "we use cookies".

The machine-readable future (and how to be early, cheaply)

The direction of travel is unmistakable: consent is becoming a signal exchanged between software, not a dialog box aimed at a human. Global Privacy Control already works this way for US opt-outs — a browser header your site must honor in several US states, no banner involved. The EU's Digital Omnibus proposal goes further, with a proposed GDPR Article 88b that would regulate expressing consent and refusal through automated, machine-readable signals. None of that is final law yet — but an agent-heavy web will need it, because the alternative is robots clicking Accept All forever. Being early costs little: honor GPC today (we send a server-side evidence record when we do), keep your consent state machine clean enough that a new signal type is one more input, and never build flows that only work if a human sees a modal.

The checklist

1. Check your logs for the major AI user agents — know your actual bot share before deciding anything. 2. Decide your crawler policy per class (training vs retrieval vs user-triggered), implement it in robots.txt, and document the decision with a date. 3. Treat detectable automation as consent-denied; never gate content behind the banner. 4. Audit your consent records for obvious automation before relying on them. 5. Discount consent-rate analytics for bot inflation before making spend decisions on them. 6. If you run a chatbot: label it, name the model provider in your privacy policy and subprocessor list, state the training stance. 7. Honor GPC now; architect for machine-readable consent signals generally. None of this requires predicting how the law settles. It requires noticing that the audience changed — and that the gap between what your site does and what your documents say is now being read by regulators, buyers, and machines alike. Keeping that gap at zero is the entire job, and it is automatable: that is what our monthly sweep does.