Can you influence what AI says about you?

Yes, but not the way most advice assumes. AI assistants build answers by fetching web pages at the moment a person asks, so what an assistant says about a company is assembled from the specific pages it retrieves in that moment. If the pages an assistant reads carry your facts, your facts show up in the answer. If they do not, no amount of prompt tracking changes what gets said.

The part most influence efforts miss: assistants read a much narrower set of pages than people expect, and most optimization is aimed at pages assistants never fetch.

How an AI answer is assembled

Assistants do two separate things with the web:

Training crawls ingest large archives on a slow cycle. They shape a model's general knowledge, not this week's answers.
Retrieval fetches happen live, when a person asks. The assistant runs a search, fetches a handful of pages, and composes an answer from what those pages say right now.

Retrieval is where influence is possible, because it is repeated, current, and observable. Our network covers 10,000+ sites and our logs record more than 200M AI bot fetches a day, fetches that page analytics never see, because AI bots do not execute JavaScript.

What actually gets read

Three patterns from our logs, August 2026, aggregated across the network:

Reading concentrates hard. On one marketing-related topic surface, the three most-read publications carried 41%, 16% and 13% of all AI reads, 70% of the surface from three sources. Winning presence on a small set of heavily fetched pages beats thin presence everywhere.
Reads dwarf citations. Only a small fraction of what an assistant reads is ever cited in the visible answer, on the order of 5–15%, and advanced agents often cite nothing at all. A citation report sees the tip. The reading behind it is the influence surface.
Answers send people back, a little. In 14 days, one explainer on this exact topic was read 307 times by assistants and received 48 human click-throughs from AI answers. That is an unusually tight loop. On tollbit.com over the past month the ratio of AI reads to AI referrals ran about 1,000 to 1, which is closer to normal.

What works

Publish the answer where assistants already read. That means two places: your own site, served in a form assistants can parse, and the third-party pages that already carry your topic's reads.
Facts over adjectives. Prices, dates, specifications, availability, comparisons. Assistants compose from verifiable statements and discount promotional tone.
Answer-first structure. The question as the heading, the direct answer in the first sentences, supporting detail below. Retrieval systems reward pages that resolve a query without excavation.
Consistency across surfaces. When the same fact appears identically on several pages an assistant fetches, confidence in that fact rises. Contradictory numbers across your own pages work against you.

What does not work

Measurement without publishing. Prompt-tracking dashboards describe the scoreboard. They do not move it.
Content aimed at pages AI never fetches. Effort follows instinct toward famous sites. Fetch logs frequently point somewhere else.
Promotional copy. Superlatives without verifiable facts are the content assistants are best trained to ignore.
Pages assistants cannot read. A JavaScript-rendered page hands a bot a few dozen words of app shell. Your facts are not in the answer because they were never in the fetch.

Where TollBit fits

AI is its own visitor. You already serve humans differently on desktop, mobile, in an app and by email, and the agent-facing version of your site is the surface where you control the facts assistants read. TollBit Intelligence shows the questions agents carry about your category, built from reads across 10,000+ sites. The Agent Site CMS publishes the answer page to bots without touching the human page. The same server logs then show whether the page was picked up.

Frequently asked questions

How fast do changes show up in answers? Retrieval is live, but caching varies by platform, from hours to weeks. Expect a change to propagate across assistants over days, not months, and watch your own logs for the first re-fetch.

Is serving bots a different version of a page deceptive? Not when the rule is same facts, cleaner format, bots only. The line to hold: facts identical to what a human would find on the human page, and any paid placement on a third-party page labeled and approved by that publisher.

How do you measure whether it worked? Three observable layers, in order. AI reads of the pages carrying your facts (server logs), human click-throughs from AI answers (referrer data), and presence of your facts in answers to the questions that matter to you, checked before and after.