What comes before GEO?

Readability, usability and measurement, in that order. GEO (generative engine optimization) works on how AI answers describe you, and it assumes the assistant can fetch your pages, parse them and complete tasks on them. If it cannot, optimization is measuring a closed door.

GEO and AEO, defined fairly

GEO is the practice of making a brand or its content more likely to appear, and appear accurately, in answers generated by ChatGPT, Claude, Perplexity, Gemini and the rest. AEO (answer engine optimization) is the same job with a different emphasis. A third label will probably win eventually.

The tools in this category run a chosen set of prompts against assistants, record who gets mentioned or cited, and recommend content changes. That is useful work: it shows where you stand on the questions you thought to ask and gives a team a trend line.

What it does not do is check the floor it stands on. A prompt tracker reports absence. It cannot say whether the assistant never fetched your page, fetched a JavaScript shell, or was blocked by a bot-management rule. Those are three problems with three owners, and the fix for each is infrastructure, not copy.

Layer 1: readable

The question for simple bots and crawlers is "are you readable?" Most retrieval bots fetch raw HTML and do not execute JavaScript. A client-side app returns a few dozen words of shell to that fetch, and your product facts are not in the answer because they were never in the response.

The common failures: JavaScript-only pages, product configurators, specifications shipped as PDFs, slides embedded as images, expired URLs still in the sitemap. The fix is server-rendered text, slim HTML, pre-computed answers and a sitemap bots can trust. Weight matters. A Forbes article ran 1,201 KB as HTML and 41 KB as Markdown (97% smaller). Every byte of template is a token the model pays for before it reaches your facts.

Layer 2: usable

The question for computer-use agents is "are you usable?" These agents navigate, click and fill forms on a person's behalf, and the patterns that lift human conversion (overlays, hover menus, CAPTCHAs, sticky elements, re-rendering DOMs) are the ones that trap them. In our benchmark with KERNEL across five test stores and 1,000 runs, the agent-facing version of the same site completed tasks 24–35% faster with up to 38% fewer steps and zero failed runs.

An agent that cannot finish on your site finishes on someone else's. No prompt moves that number.

Layer 3: measured

The question for you is "can you see them?" Google Analytics never records an AI visit, because the tag never runs. Server logs do. The median North American site saw 1 AI bot visit per 93 human visits in H1 2026 (1 per 33 in Europe), and on tollbit.com over the past month AI reads outnumbered AI referrals by about 1,000 to 1. Until you can see reads by agent and page, you cannot tell whether an optimization was read, ignored or blocked.

Measurement also says what to optimize for. Reads across 10,000+ sites reconstruct the questions agents carry about a category, which routinely differ from the prompt lists teams draft by hand.

The order of operations

  1. Readable. Serve identified AI agents a server-rendered, slim version of every page. Fix the PDFs, the configurators and the sitemap.
  2. Usable. Give advanced agents a surface without overlays and unstable navigation, so tasks complete.
  3. Measured. Forward server logs. Count reads by agent and page, blocked fetches, and referrals.
  4. Then optimize. Publish answer-first pages where the reads say the questions are, and check the answers with whichever prompt tracker you prefer.

AI is its own visitor, and it needs its own experience before it can be optimized for. TollBit sits in layers 1 to 3: the agent site that serves bots a readable, usable version of your domain, and the analytics and network intelligence that measure what they read. GEO starts where that leaves off, and works better on an open door.

Frequently asked questions

Do I still need GEO if I have an agent site? They are different layers. An agent site makes your facts reachable and your tasks completable, and the logs show what was read. A prompt tracker checks what the answers said. Most programs end up with both.

Does GEO work on a JavaScript-rendered site? Only on whatever the bots can read, which on a client-side app is usually the title and a few dozen words. Fix rendering for bots first, or the optimization has nothing to work on.

How do I know which layer I am missing? Fetch your key pages with curl as an AI user agent and count the readable words. Run a five-task test with a computer-use agent. Pull a week of server logs and count AI reads by status code. One test per layer.

Related