Does ranking in Google get you cited by AI? We tested it.

What the numbers mean

Why does not ChatGPT mention my company even though I rank?

Because getting cited by AI is two gates, not one. The first gate is retrieval: can a search index surface your page as a candidate at all? Ranking answers that gate fairly well, a live web index surfaced the tested pages on 27 of 45 queries (60 percent). The second gate is lift: once your page is in the retrieved pool, does the model judge it the clearest, most self-contained answer worth quoting? That is where the tested pages lost, cited on just 7 of 45 (15.6 percent).

The pattern inside the misses is the useful part. On several queries the page ranked higher than the competitor the model chose to cite: pages at positions 2.2, 2.7, 3.1 and 3.2 went uncited while pages ranked lower were quoted. Ranking above someone did not win the citation from them. The pages that lost tended to phrase their key numbers as a vendor price rather than as a neutral market fact, which reads less like a quotable answer and more like a pitch.

Does SEO help with AI answers?

For the first gate, yes, and this study shows it: the pages ranked, so they were retrievable, surfaced on 27 of 45 queries. SEO is what makes AI visibility possible. But SEO stops there. It converted to a browsing citation on only 7 of 45, and in a separate 9-query check it did nothing at all for a non-browsing model, which answered from training memory and named the pages 0 times. Treating on-page SEO as an AI citation strategy overstates what it buys.

Why did a non-browsing model never mention the pages?

A model with no web access answers from what it absorbed during training, which is shaped by how often and how widely a source is mentioned across the web over time, not by a single well-ranked page. One strong blog post does not move that memory. The 0-of-9 result from the no-browsing sub-test is the recall gap: the pages have not accumulated enough authentic, third-party mention density to be remembered when the model cannot look anything up.

So what actually moves AI citation?

Two motions, matched to the two gates. To win the lift gate, make the page the single clearest answer: a direct answer in the first hundred words, phrased as a neutral fact with specific numbers before any vendor framing, question-shaped headings, and real self-contained stat sentences a model can lift whole. To close the recall gap, earn authentic mentions over time, genuine roundup inclusions, real reviews, and being named in the threads that answer the buying question, never manufactured links.

Plain-English methodology

How we tested it: sample, engines, and dates

Question

For pages that already rank on Google for a query, does an AI assistant actually cite them when a person asks that same question?

Sample and dates

The main run used the 45 highest-impression queries the site ranks for at Google position 12 or better over the prior 90 days, taken from Google Search Console. Selecting by ranking and impression, rather than by topic, keeps the sample from being curated to flatter the result; every qualifying query is included verbatim, including odd, long, or branded ones. All observations were collected in August 2026. This is a first-hand study of a single site, not a random sample of the web.

Engines tested

• Live web-search index (Brave), a proxy for the retrieval layer: does the page get surfaced at all? Run across all 45 queries.
• Browsing AI model (gpt-4o-search-preview), which fetches live sources and cites them: does the page get named in the answer? Run across all 45 queries.
• Non-browsing models (gpt-4o-mini and Grok-3), which answer only from training memory: does the model recall the page without searching? Run as a separate, smaller 9-query sub-test, so its 0-of-9 result has a denominator of 9, not 45.

What was measured

For each of the 45 queries we recorded two binary outcomes: whether the web index surfaced the page, and whether the browsing model cited it. Retrieval and citation are independent, so some queries were cited without being surfaced by the web index, and many were surfaced without being cited. The non-browsing recall check was scored separately on its 9 queries.

How the measurement was made

1. Each query was issued as a natural user question, not as a keyword.
2. Queries were run through model APIs rather than a signed-in browser, so the model could not see who was asking and parrot it back. This avoids a common bias where a personal session already knows your company.
3. A citation was counted only when the browsing model named or linked the page in its answer. For non-browsing models, a mention of the company or page by name counted.

Interpretation and limitations

• This describes 45 queries on one site and should not be projected to all pages or all AI engines.
• A live web-search index is a proxy for retrieval, not the browsing model's own retriever, so the two are measured with different tools.
• API probing understates real-interface behaviour; a stable share-of-voice number needs many runs per engine, and this is a single pass.
• Citation behaviour changes as models and their retrieval stacks update, so a repeat of this test on another date could differ.