Owned pages only win the empty slots

An empty slot is an AI answer where no source has yet given the engine a reason to name your brand, and the 2026 experiment data says owned pages can fill those slots but rarely win the contested ones.

That distinction explains why two well-run tests reached results that look contradictory at first. One found a brand's own pages driving most of its new AI mentions. The other found a brand's own pages barely registering next to third-party coverage. Both are plausible once you ask how much competing evidence the engine already had for each query.

#Experiment one: Ahrefs promoting itself

Ahrefs published 34 self-promotional pages across 5 domains, promoting two things: its Brand Radar product and its Evolve conference. It then tracked 9,886 answers from ChatGPT, Gemini, Perplexity, and Copilot between February 7 and May 31, 2026 (Ahrefs).

The two subjects behaved very differently.

For the conference, 82% of new mentions in previously empty slots came from answers that cited Ahrefs' own pages (Ahrefs). The owned content did the work.

For the product, only 6% of new mentions came from answers that cited Ahrefs' pages. The other 94% came from third-party content (Ahrefs).

Same company, same tactic, same tracking window. The gap is the category. A specific conference has few independent sources describing it, so the organizer's page is often the best evidence available. An SEO tool sits in a category with years of reviews, comparisons, and forum threads, and the engine leans on those.

The query wording showed the same pattern inside one subject. Evolve appeared in 66.4% of responses to "best SEO conferences 2026" but only 15.8% of responses to "best marketing conferences 2026" (Ahrefs). The narrower query has fewer credible candidates. The broader one has more established events competing for the same handful of names.

#Experiment two: a brand starting from zero

Search Engine Land published two GEO experiments on September 14, 2026 (Search Engine Land). The second is the cleaner test for this question: a SaaS link-building agency with no baseline AI visibility, tracked for 30 days across 15 commercial-intent keywords on six AI platforms, with 437 citation events logged (Scalevise summary).

Third-party listicles produced 85.8% of source mentions. The brand's own listicle produced 14.0%. PR produced 0.2% (Scalevise summary).

Link building for SaaS is a crowded service category. Plenty of agencies, plenty of roundups. A new entrant's own "best agencies" page competes against every existing list, and the engine had no prior reason to trust it. Third-party lists gave the engine an independent source that already sat in the retrieval set.

The first Search Engine Land experiment, a consultant brand tracked over several months across ChatGPT, Claude, Gemini, and Perplexity with 775 citation events, also found listicles the most effective format (Scalevise summary). The authors caution that the results "should not be read as a universal ranking rule for every query or AI system," which is the right reading.

#Cited is still not recommended

Owned pages carry a second risk even where they get cited. In the Ahrefs data, 43% of the cases where the engine cited a conference page never mentioned Ahrefs Evolve in the answer. For product pages the figure was 11% (Ahrefs).

A larger audit of Google AI Overviews found the same failure at a higher rate: self-promotional listicles were used as a source while the answer recommended a different brand in 69% of citations (Search Engine Land). We covered that study in your own listicle recommends your rivals. A citation in the source strip is a weaker outcome than a name in the sentence.

#Neither kind of placement stays put

Both studies found that citations decay quickly. In the Search Engine Land experiments, roughly half of cited sources disappeared within 30 days (Scalevise summary). In the Ahrefs test, citations appeared on only about one in three eligible days between a page's first and last appearance, and nearly a quarter of cases were one-time citations that never recurred (Ahrefs).

That changes how a placement should be valued. An earned listicle mention is not an asset you bank. It is a source the engine may or may not retrieve next week, so a single win in a weekly report says little.

#How to sort your queries

The practical move is to classify each tracked query by how contested it is before choosing a tactic.

1. List the brands each engine already names. Run the query several times per engine and record every brand mentioned. Few names, and inconsistent ones, suggest an open slot.
2. List the sources behind those names. If the cited sources are mostly vendor pages, owned content has a fair chance. If they are review sites, roundups, and forums, the engine has settled on third-party evidence.
3. Match the tactic to the slot. For narrow, specific queries (a named event, a niche use case, a product attribute nobody else documents), publish the clearest owned page on the subject. For broad category queries, spend on getting into the lists and reviews the engine already retrieves.
4. Measure the mention, not the citation. Track whether your brand appears in the answer text, separately from whether your URL appears as a source.
5. Re-measure monthly. Given the decay rates above, a query you won in one month needs checking again the next.

#What this does not show

Both experiments are single-company tests run by marketers with an interest in the result, on a small keyword set. Neither isolates the category effect directly; the contested versus empty reading is our interpretation of why the two Ahrefs subjects diverged, supported by the query-wording gap. The numbers are a reason to classify your own queries, not a benchmark to expect.

The working rule holds up across both datasets. Owned pages are the right tool where the engine has nothing better to cite. Where it already has a shelf of independent sources, your page is one more vendor claim, and the shelf wins.