Agent visibility: your second audience is software

For twenty-five years a website had one job: convince a person. In 2026 it picked up a second visitor — software that fetches, parses and decides on someone's behalf. For the first time in the web's history, more than half of its traffic isn't human.

The short version: agent visibility is the work of being readable and recommendable to the software agents acting for your buyers. It's one layer above AI visibility — there you're trying to be *named in an answer*; here you're making the machine able to understand, trust and act on your business. It's early. That's the point.

The web quietly changed hands

• For the first time, non-human traffic passed 50%: Cloudflare put bots at 57.5% of HTML web traffic against 42.5% human (Cloudflare); Imperva's Bad Bot Report puts automated traffic at 53% of all web traffic, up from 51% the year before (data reference).
• The growth is agents: daily requests from AI agents on Cloudflare's network grew more than 1,700% in a year (Cloudflare).
• On one developer-docs platform, agents became the majority reader — 66% of traffic in August 2026 (Mintlify).
• Real brands already feel it: at the UK retailer John Lewis, agentic searches went from 0.3% to 2.5% of visits in a year (via Digiday).
• One caveat before you panic: "over half" counts *requests* — crawlers, API calls, fetches — not human *visitors*. Your analytics still show people. The point isn't that your audience shrank; it's that a new one appeared.

You now have two audiences

• People browse: they wander, judge on feel, decide slowly.
• Agents fetch: they arrive with a task, read your data, and act. No patience for your hero animation, no opinion on your typeface.
• The agent doesn't evaluate your *design* — it evaluates your legibility. Can it find the fact, trust it, and use it? That's a different brief from the one most sites were built for.

Not all bots are equal — training vs search

The number that matters isn't how many bots crawl you; it's *why*. Of AI crawler requests in May 2026: 51.8% training, 35.7% mixed, and just 9.3% search (Cloudflare). And the returns are wildly uneven — *pages crawled per single visit sent back*:

• ClaudeBot ~23,951:1 and GPTBot 1,276:1 — training crawlers. They take a lot and return almost nothing.
• PerplexityBot 111:1 and Copilot 33:1 — answer crawlers. They cite, so they send traffic.
• Google 4.9:1 and DuckDuckGo 1.5:1 — search. Crawl and referral stay roughly in balance (SEOmator via Cloudflare Radar).

So "should I allow bots?" is the wrong question. The right one is which purpose: extraction, or answer?

Should you block the training bots? The honest answer

• If your product is the content — a publisher, a newsroom — training crawling is extraction: they take the thing you sell and send back nothing. Block or license it. It's why GPTBot is the single most-blocked crawler (data reference).
• If your product is a brand — a studio, a service business — training is closer to free memory. Being in the weights is being remembered; when a buyer later asks "who's good at this," the model reaching for your name is the payoff. You can't measure it like a click, but it's the same instinct as wanting a Wikipedia page.
• The trap: Google merged its search crawler and its AI trainer (Google-Extended) into one user-agent — so blocking Google's AI training also forfeits Google search (Cloudflare).
• Rule of thumb: keep the answer/search crawlers open, always — that's how you get named. Keep training open for brands; make it a deliberate per-client call for publishers. Never blanket-block.

What agent visibility actually involves

• Agent-readability — a machine-first surface: clean semantics and schema, machine-readable pricing and availability, markdown-first content, and an agent/llms manifest. Where it fits, an MCP endpoint so agents query live facts instead of scraping a guess.
• Allow/block policy — a *deliberate* per-crawler decision (training vs search), written down and revisited as the crawlers change.
• Measurement — agent traffic in GA4 and server logs, crawl-to-refer ratios per crawler, AI-referral conversion.
• Presence where agents read — structured directories, reviews and comparison data: the same third-party proof, tuned for machines instead of people.

Why being early matters

• AI referral is still small but it's the fastest-growing channel, and it converts: AI-referred retail traffic grew +393% year over year, converting 42% better — retail-specific, but the direction is clear (Adobe via data reference).
• The field is pre-consensus: most teams haven't decided a crawler policy, and almost nobody has built an agent-readable surface. Categories get named early — and the studio that *writes the definitive guide* is the studio that gets *named in the guide*.

What to do now

• Decide your training vs search crawler policy — on purpose, not by default.
• Make the facts agents need machine-readable (schema, pricing, availability).
• Instrument agent traffic (GA4 + server logs) and your crawl-to-refer ratio.
• Publish the definitive page for your category — the one an agent (or a person) will lift the answer from.
• Watch the trend. The numbers move monthly; the structure doesn't.

The second audience doesn't browse. It reads, decides, and moves on. The businesses that win the next decade will be the ones whose facts a machine can find, trust, and repeat — while everyone else is still optimising for a scroll that's quietly ending.

Get in touch: thecollective@memeador.com