What is share of reads, and how is it different from share of voice?
Share of reads is your brand's portion of the AI reads on a topic: of all the pages AI agents fetched about a category, how many carried you. It is observed in server logs across 10,000+ sites. Share of voice is your portion of the mentions in answers to a prompt list someone chose to run.
One measures what AI reads to form an answer. The other measures what a sample of answers said. They diverge because assistants cite on the order of 5–15% of what they read, and advanced agents often cite nothing.
How each is measured
Share of reads starts with a fetch. An AI agent requests a page, the server logs it, and the page is classified by topic and by the brands on it. An AI read is the agent-web equivalent of an impression: one agent, one page, one fetch. Add them up by brand across a topic and divide. Because bots do not execute JavaScript, this exists only in server logs.
Share of voice starts with a prompt. A tool runs a list of questions against ChatGPT, Claude, Perplexity and others on a schedule, parses the responses, and counts which brands were named. Divide your mentions by all mentions.
What share of reads can tell you
- Where the demand is. The questions agents carry about your category, reconstructed from what they fetched, including ones nobody put on a prompt list.
- Where it concentrates. On one topic surface in August 2026, three publications carried 41%, 16% and 13% of all AI reads. Those are the pages your facts need to be on.
- Who is reading. Reads split by agent, so an indexer's sweep and a retrieval bot's per-question fetch show up as different things.
- Whether a change was picked up. Publish a page, watch the fetches arrive in the same logs, within a recrawl window that varies by platform from hours to weeks.
What it cannot tell you: what the assistant wrote. A read is a page in the context window, not a sentence in the answer.
What share of voice can tell you
- What a sampled answer actually said. The closest thing to what a customer sees.
- Tone and framing. Whether you were recommended, listed or warned against.
- A legible trend line for a board, on a fixed question set.
What it cannot tell you: anything about questions not on the list, the 85–95% of reading that produced no citation, or which pages to change to move the number. The prompt list is a hypothesis about demand. Reads are the demand.
Side by side
| Share of reads | Share of voice | |
|---|---|---|
| Source | Server logs, observed fetches | Sampled prompts, parsed answers |
| Unit | One agent fetch of one page carrying the brand | One mention in one sampled answer |
| Coverage | Every question that touched measured sites | Exactly the prompt list |
| Tells you what to do | Which pages and publications to be on | That you are up or down |
| Blind spot | Sites outside the measured network | Everything the assistant read but did not cite |
Why reads lead
The chain runs in one direction: a page is published, an agent fetches it, an answer is composed, a person maybe clicks. Reads move first, within the recrawl window. Mentions move when enough fetched pages carry the fact. Referrals trail both, at hundreds or thousands to one (about 1,000 to 1 on tollbit.com over the past month, 227 to 1 across European sites in Q2 2026).
Share of reads is the number that changes when you publish. Share of voice confirms it did, later, on the questions you remembered to ask.
TollBit Intelligence reports share of reads by topic, brand, agent and publication across our network. It sits before any prompt tracker and pairs with whichever one you run.
Frequently asked questions
Is an AI read the same as an impression? Close enough to use the word. One AI agent fetched one page, once. Unlike a display impression, the agent read the whole page into a context window, and a human may never see it.
Does a read mean my brand was mentioned in the answer? No. It means a page carrying your brand was in the material the assistant worked from. Whether the assistant named you depends on what the page said and what the person asked.
How complete is the read data? It covers the sites being measured. Our network is 10,000+ sites, large but not the web, and no vendor has the whole web. Ask any provider what they observe directly and what they infer.