AI already knows who you are. It will not volunteer you.

Ask an AI assistant about a specific business by name and it will almost always answer accurately. Ask the same assistant a category question that the business should win, and the business usually does not appear at all. On our own week-1 scan (2026-08-06), the five major assistants mentioned The AI Citation Institute in 100% of branded answers (15 of 15, all five engines at 100%). On the same day, on the same instrument, we appeared in 0% of discovery answers: 0 of 130 on the buyer panel, 0 of 90 on the seller panel, 0 of 25 on the core panel. The assistants could describe us, cite our pages, and summarize our pricing. Not one of them named us when a buyer asked the category question we exist to answer. We see the same split, in varying degrees, across every business we measure. Identity is nearly solved. Discovery is the entire game.

What did we measure, and how?

We run frozen query panels: fixed lists of questions that do not change from week to week, so results are comparable across time. Each panel has two kinds of queries. Branded queries ask about the business by name, for example “What is Gavelist?” or “Is The AI Citation Institute legitimate?” These test whether the assistant knows the business exists and can describe it accurately. Discovery queries are category questions where the business should be a candidate answer, for example “What software should an auction house use to catalog lots?” These test whether the assistant will volunteer the business to someone who has never heard of it.

We run every panel against five engines: ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews. High-intent queries are sampled three times per engine and scored by majority vote, because assistant answers vary between runs. Scans run on a weekly cadence. A mention is counted when the answer names the business; a rate is the share of answers that do. The full instrument is documented on our methodology page, and the running numbers, including our own, are public on our scoreboard.

What do the numbers show?

Across all nine subjects, eight read branded at 66.7% or higher, and six read 100%. Discovery rates over the eight measured subjects ranged from 0% to 66.7%. One subject, a community/benefit auction organization, runs two properties on a single domain: its main property reads 66.7% branded and 17.4% discovery, while a second property on the same domain reads 0% branded and 3.3% discovery, because the engines do not yet associate that second URL with the business. We count it as one business and report the main property's reading, with the second property disclosed. The ninth subject, the auction-realty firm, read 0% branded; we attribute that to an unresolved identity-curation issue in our own roster, and we excluded its discovery reading from the summary rather than let a defect on our side pose as a finding. It is disclosed in full in the limitations section below. An earlier version of this post stated nine subjects, then eight; both counts were wrong. The table above is the complete enumeration and the counts are computed from it.

The gap is not a small-business artifact. It appears at every size we measure, and it moves. The online auction platform read 2.8% discovery on an earlier platform-set scan (2026-07-14) and 41.7% on 2026-08-02. Gavelist's weekly core-panel mention rate rose from 31.3% (week of 2026-04-27) to 65.3% (week of 2026-06-22), and reads 59.5% at the most recent completed scan (2026-08-09; 48 queries, five engines). The panel itself has widened from 20 to 48 queries over the engagement as new buyer questions were discovered, so readings across the full span are directional, not point-comparable. Discovery rates change week to week. Branded rates, once established, sit at or near 100% and stay there.

Correction, 2026-08-10: an earlier version of the paragraph above dated the 31.3% and 65.3% readings to 2026-05-23 and 2026-06-25 — the transcribed weekly series places them in the weeks of 2026-04-27 and 2026-06-22 — and presented the 53.0% run of 2026-07-23 as the most recent reading. A later paragraph also described this series as measured on an unchanged panel; the panel in fact widened from 20 to 48 queries across the window, which makes part of every downswing dilution rather than decline. The current figures are from the completed run of 2026-08-09.

What happened when we measured ourselves?

Our own case is the cleanest arc in the data, because we controlled the launch date and we publish our own numbers. On 2026-07-19, before our site launched publicly, our branded rate was 80% (12 of 15 answers). The assistants already partially knew the organization from indexed traces. Our discovery rate on the vendor-selection panel was 0% (0 of 128 answers). The site launched publicly on 2026-07-31.

On 2026-08-06, seven days after launch, the branded rate was 100% (15 of 15), with all five engines at 100%. The branded answers were not vague. On Gemini and Google AI Overviews, they cited our pages 13 times across 7 distinct URLs, including pricing, methodology, and blog pages. One engine lagged: Claude's web search still reported it could not find the organization in the pre-backfill scan of that same week, which is a useful reminder that engines do not update in lockstep.

Discovery did not move at all. Buyer panel: 0% (0 of 130). Seller panel: 0% (0 of 90). Core panel: 0% (0 of 25). Meanwhile the incumbent tools in our category appeared constantly in those same 130 buyer answers: Semrush in 71.5%, Profound in 60.8%, Otterly.AI in 60.8%.

Within one week, five AI systems went from partial recognition to complete, citation-backed knowledge of who we are. None of them will yet volunteer us to a buyer.

That is a day-0 reading for a new site, and it is exactly what our model predicts. The series continues weekly and we will publish the discovery curve as it develops, whatever it shows.

Why does this happen? (Our hypothesis, honestly labeled)

This section is our interpretation. The data above establishes the split; the mechanism is inference. We think branded and discovery answers are produced by different processes with different inputs. Identity appears to be established by consistent entity signals and ordinary indexing. When a query contains the business name, the engine's retrieval step can find the business's own site, its structured data, and third-party records of it, and the answer follows. This is why a one-week-old site with consistent signals can reach 100% branded across five engines: the question hands the engine the answer's address.

Discovery is different. When a buyer asks a category question, the engine is not retrieving the business; it is retrieving sources that answer the category question, and those sources are overwhelmingly listicles, comparison pages, community threads, and category roundups. A business appears in discovery answers roughly to the extent that it appears in that corroborating layer. Its own site, however complete, is one voice saying “we are a good answer,” and engines discount self-description for exactly the reason a sensible person would. In our category, the tools appearing in 60.8% to 71.5% of buyer answers are the tools that saturate the comparison-content layer.

The stakes of this mechanism are rising with usage. Pew Research Center found that 49% of U.S. adults reported using AI chatbots in a survey fielded February 17-23, 2026, and Edison Research (with SSRS) reported in March 2026 that 52% of Americans use AI platforms every week. On the search side, Pew measured AI summaries appearing in about 18% of searches in its March 2025 panel study of 900 U.S. adults' browsing, while BrightEdge tracked AI Overviews present on about 48% of tracked queries by February 2026. And when an AI Overview is present, Ahrefs measured pages ranking #1 taking a 34.5% lower click-through rate in early 2025, measured at 58% lower on December 2025 data. If the AI layer increasingly intermediates the category question, then the discovery rate, not the branded rate, is the number that determines whether new customers ever encounter a business. Sources are listed in full at the end of this post.

What should a business owner actually do with this?

First, stop spending effort proving to AI systems that you exist. Across the nine subjects we measured, six read 100% branded and two more read 66.7%. If your branded reading is healthy, further investment in being describable has little room to pay off.

Second, check your own split before assuming it. Ask each major assistant about your business by name, then ask the category questions your customers actually ask, without your name in them. The pattern in the table above took an instrument to quantify, but the direction of your own split is checkable in an afternoon. As disclosure: measurement is what our organization sells, and we offer a free version of this check, but the manual version costs nothing and tells you most of what you need.

Third, if your branded reading is broken (an assistant confuses you with another business, or cannot find you), fix that first. It is the cheap problem: consistent name, address, and description across your site and the major records of your business. The auction-realty firm in our data is a live example of how an identity defect zeroes out everything downstream.

Fourth, direct discovery effort at the corroborating layer, not at your own site. The question to ask about any marketing activity is whether it produces an appearance in the material engines cite when answering your category question: independent comparison pages, industry lists, community discussion, editorial coverage. Your own site establishes that you exist. Other people's pages establish that you are an answer.

Fifth, measure on a cadence, because single readings mislead. On the frozen panel — the original 20 queries on the two probes with zero instrument events across the record, where nothing changed but the answers — Gavelist read 67.5% in the week of 2026-06-01, 60.0% two weeks later, and 80.0% by the week of 2026-07-06. The growing all-engine panel swings harder still, and part of those swings is the exam widening from 20 to 48 queries, not visibility moving. A one-time reading during any of those swings would have told a false story in one direction or the other.

How this could be wrong

Measurement surface: we measure engines through their API surfaces, not the consumer apps. Consumer-surface answers can differ in retrieval behavior and formatting, and our rates should be read as API-surface rates. Asymmetric detection: branded queries contain the business name, so detecting a mention is trivially reliable, while discovery answers require recognizing the business among free text; this asymmetry could overstate the gap at the margin, though it cannot explain readings of 0 of 130.

Panel sizes: some panels are small. Our branded panel is 15 answers per scan and our core discovery panel is 25. Rates on small panels move in coarse steps, which is why we report every rate as n of m where we have the counts. Engine drift: engines change weekly, so a rate measured 2026-08-02 is a fact about that week, not a stable property of the engine. And the excluded subject: one auction-realty firm read 0% branded on 2026-08-02, which we attribute to an unresolved identity-curation issue on our side of the instrument, not to the engines. We excluded it from the summary and disclose it here rather than silently drop it.

Where this series goes

Everything above was measured on our frozen panels between 2026-07-19 and 2026-08-06: our own baseline on 2026-07-19, the platform rebaseline on 2026-07-24, client seller-panel scans on 2026-08-02, and our week-1 scan on 2026-08-06. The site whose discovery curve we most want to watch is our own, because it started at a clean zero on a known launch date. The series continues weekly, and the running numbers, including ours, are public on the scoreboard.

Sources

1. Pew Research Center — Americans and AI 2026 (ATP Wave 187, fielded Feb 17-23, 2026)
2. Edison Research with SSRS — weekly AI usage, March 2026
3. Pew Research Center — Google users and AI summaries (March 2025 panel study)
4. BrightEdge — AI Overviews at the One-Year Mark (Feb 2025-Feb 2026)
5. Ahrefs — AI Overviews reduce clicks (updated on December 2025 data)

Every external figure above was verified against the primary source before publication. Our own figures come from the instrument described on the methodology page; the claims ledger for this post is part of our research record.