How people search on AI assistants
How people search on AI assistants is a core concept. Searching on AI assistants is the shift where a buyer asks an AI "who is best at X" instead of scrolling ten blue links. The assistant returns one answer that names only a few businesses and omits the rest. The visit it sends often arrives with no referrer, so it lands in GA4 as "Direct" rather than as an identifiable AI referral. Most tools stop at telling you AI answers exist and scoring your visibility. Ignite's point is why that is not enough. The conversation happens off your property, the referral is undercounted, and the brand named in the answer is chosen before the buyer ever reaches a website. Ignite measures that gap today. It splits branded questions from non-branded and grounded answers from parametric, and it is building the loop that applies on-site fixes and proves the answer moved.
What changes when a buyer asks an AI instead of searching?
The selection moves inside the model. When a buyer asks an AI assistant "who is best at X," the assistant collapses the old ten-link shortlist into one spoken answer. It names a few businesses and omits the rest. There is no page of links to scroll and no obvious second place. The buyer often acts on the named options without ever seeing the sources the model drew from.
For two decades the front door to a business was a search engine. A buyer typed a query, saw ten blue links, and built their own shortlist. That changes when the buyer asks an AI instead. A business that would have ranked eighth on a results page still had a chance to be clicked. A business the assistant does not name is simply absent from the buyer's consideration, and it gets no signal that the conversation happened.
In pilot measurement a reference brand is routinely absent from the non-branded category answers about its own field, named in none of them across a sampled set even though the brand exists and has a website. (Illustrative of the pattern, not a benchmark.)
In pilot measurement a reference brand is routinely absent from non-branded category answers about its own field. (Illustrative of the pattern, not a benchmark.)
Why does the AI answer name only a few businesses?
Because the assistant returns one synthesized answer instead of a ranked list, the field of consideration shrinks to the handful of names the model volunteers. Being present somewhere on the open web is not the same as being named in the answer. In pilot measurement a reference brand is routinely absent from the non-branded category answers about its own field, even though the brand exists and has a website. (Illustrative of the pattern.)
Competitors frequently occupy that named space instead. In pilot measurement rivals in aggregate are cited roughly 2x more often than the brand across the category questions measured. That is the shape of the problem: the assistant is already recommending someone in your category, and it may not be you. (Illustrative of the pattern, not a prediction of any reader's result.)
In pilot measurement competitors in aggregate out-cite a brand by roughly 2x and the brand is routinely absent from non-branded category answers. (Illustrative of the pattern, not a benchmark.)
Why do AI referrals show up as "Direct" in GA4?
AI referrals show up as "Direct" because many AI-assistant visits reach a site unattributed, so they fall into GA4's "Direct" channel instead of an AI-identified one. GA4 does expose a native "AI Assistant" channel. But so many AI-driven visits never announce where they came from that the channel is a floor, never a total.
The gap is visible in the raw numbers. In pilot analytics the great majority of sessions classified as "Direct" dwarf the handful tagged as coming from an "AI Assistant" over the same period. The true count of AI-influenced visits sits above the tagged number, hidden inside Direct. No amount of channel tuning fully recovers it. Any AI-referral figure a business reads off GA4 should be treated as a lower bound. (Illustrative of the pattern.)
Ignite's read: many AI referrals reach a site unattributed and land in GA4 as "Direct," so the GA4 "AI Assistant" channel is a floor, never a total.
In pilot analytics the great majority of AI-influenced sessions land in Direct with only a handful tagged AI Assistant over the same window. (Illustrative of the pattern.)
Why does a brand never see the AI conversation about it?
A brand never sees the AI conversation because the exchange happens inside a chat interface the business is not part of. The buyer asks their question, the assistant answers, and the brand whose reputation is being described is not in the room. It gets no query log, no lost-click record, and no notification that a rival was recommended in its place. In pilot analytics nearly every AI-influenced visit reaches the business unattributed, logged as Direct rather than AI Assistant. Every buyer asking that question this month hears the answer, and the business never sees it. (Illustrative of the pattern.)
This is why the problem is not mainly a traffic problem that shows up in a dashboard. The recommendation, the omission, and any factual error about the business all happen before a page view could be recorded. Much of the resulting traffic is miscounted when it does occur. Seeing the conversation at all means sampling the assistants directly, not waiting for their referrals to appear in analytics.
In pilot analytics nearly every AI-influenced visit reaches the business unattributed, logged as Direct rather than AI Assistant. (Illustrative of the pattern.)
How do you find out what AI assistants say about your business?
You sample the assistants directly, asking them the buyer's questions and recording the answers. Ignite samples five engines. ChatGPT, Claude, and Gemini answer parametrically from the model's own memory with no live retrieval. Perplexity and Grok are grounded, retrieving and citing live sources. Google's AI Overview is cross-checked separately for presence rather than swept like a chat engine, and Microsoft Copilot is not directly sampled. These are never blended. Branded questions are split from non-branded ones, and grounded answers are kept apart from parametric ones, because they behave differently and averaging them would hide the truth.
Honesty rails bound every number. Ignite requires repeated samples over a rolling window before publishing any public figure, and it reports a confidence interval or withholds the number rather than showing a fabricated zero. External research supports the direction of the on-site work. Aggarwal and colleagues found that adding statistics, citations, and quotations to visible content lifted a page's presence in generated answers by up to 40%, with best-method gains of 22% in Position-Adjusted Word Count and 37% in Subjective Impression [1]. That result is measured as Position-Adjusted Word Count, how much of a page's text an answer quotes, not website traffic. It is their finding, not an Ignite promise of a reader's outcome.
Aggarwal et al., "GEO: Generative Engine Optimization," KDD 2024: statistics, citations, and quotations in visible content raised answer visibility by up to 40%, with best-method gains of 22% in Position-Adjusted Word Count (share of a page's text an answer quotes, not traffic) and 37% in Subjective Impression.[1]