The #1 AI ranking factor? What OppAlerts found, and what 34,960 AI answers add
Ben Wills found a remarkably consistent relationship between a brand's offsite relevance and its AI ranking. I think that is useful. I also think marketers need to know what it does not explain.
Ben Wills' new OppAlerts analysis offers an appealing answer to the question every AI-search marketer is asking: what gets a brand recommended by ChatGPT and Gemini? Across 100 industries, he finds that highly recommended brands tend to be surrounded by links and brand mentions from pages that are semantically relevant to their business.
It is a strong reason to investigate relevant digital PR and third-party coverage. But an offsite profile describes what the wider web says about a brand. It cannot, by itself, tell us which source an AI assistant sees, or why it picks a brand for a particular user question.
Our sixth research paper, based on 34,960 GPT and Gemini monitoring observations across 75 projects, looks at that missing step. Put the studies together and a more useful model emerges: offsite relevance helps describe a brand's footprint; onsite content and live retrieval tell us whether the evidence is available when it matters.
What the OppAlerts study actually tested
Wills queried 18 AI models and channels across 100 industries. For each industry, he used 25 variations of broad recommendation questions such as asking for the best hotel brands. He combined how frequently brands appeared with their positions in the answers.
He then looked up the third-party pages associated with more than 100,000 recommended sites. They were divided into three groups: pages linking to the site, pages mentioning the brand or domain, and pages mentioning the brand without linking. An embedding score estimated how closely those pages related to industry categories from four taxonomies.
His central observation is that the third-party pages around highly recommended brands tend to be about the same industry. He reports that the category-relevance pattern held across 21,600 curves generated from different industries, channels, taxonomies and evidence types. Unlinked mentions sometimes showed an even stronger pattern. That is interesting, especially for PR teams whose work rarely results in a clickable backlink.
Do other websites talk about this hotel in the right context?
Suppose an imaginary hotel wants to be associated with travel and hospitality.
The study found that frequently recommended brands tend to have topically relevant offsite footprints.
What I agree with
Topical context is more informative than a raw link count. A hotel mentioned by travel writers and local guides probably has a more relevant public footprint than one discussed mainly in unrelated contexts. It is reasonable to measure linked and unlinked coverage, and to ask whether the coverage matches the business and the buyer's interests.
This also fits how we work with clients at Aiso. Third-party pages can explain and substantiate a brand in ways its own website cannot. Our sixth paper explicitly acknowledges this: a brand can appear without a citation to its own domain, and offsite sources may be part of the explanation.
Where I would be more cautious
1. Consistent correlation is not causation
Wills argues that the correlation is sufficiently strong to call it causal. I would not make that leap. Well-known brands attract more topical coverage, have better-known products and are more likely to be recommended. Existing brand recognition could drive both the offsite footprint and the AI recommendation, even if no particular link had an effect.
To claim that adding relevant mentions causes more AI recommendations, we would need an intervention, or a design that isolates changes in coverage from the many other ways brands differ. Comparing brands with matched demand, size, history and existing visibility would also help.
2. Twenty-one thousand curves are not twenty-one thousand independent tests
The same industries, brands, outputs and source corpus reappear across models, categories and three overlapping evidence definitions. Multiple curves can confirm that a descriptive pattern is robust to measurement choices, but they are not 21,600 independent replications. Nor does a falling curve of relevance across categories, especially when categories are arranged from closer to farther away, establish that backlinks caused higher rank within an industry.
What I would want to see next is a clean within-industry comparison of highly versus rarely recommended brands, with effect sizes, uncertainty estimates and meaningful controls. Can the same offsite-relevance measure predict a change in ranking on questions it has not seen before?
3. Broad category rankings miss a real buyer's request
"What are the 25 best hotels?" gives a useful general picture of brand familiarity. But a buyer might ask for a quiet hotel within walking distance of a station, with connecting rooms and check-in after midnight. Our second paper found that the final prompt in a conversation often contains only part of what the user previously said. Our third paper found that restoring earlier context materially changed answers in 44.7% of paired cases.
That is why we analyze the entire request and the searches generated behind it, not just the industry a brand belongs to.
4. Similarity of external pages does not reveal the model's mechanism
Embeddings are a useful measurement tool. They do not prove the engine uses those exact embeddings, a particular category taxonomy or a single hidden "topical backlink factor" to select the answer. Training, tool-driven retrieval and answer generation are different processes. A live assistant might cite the hotel's website, use a travel review, rely on earlier brand familiarity or combine these sources. Wills' measurement does not show which occurred in a particular run.
Our complementary angle: what happens between the question and the recommendation?
For this we need to follow the request. Does the brand have content that actually answers it? Does the AI engine search? What follow-up queries does it generate? Is the brand's website or a relevant third-party page surfaced? Finally, does the answer mention the brand?
This is the model in our paper From Prompt to Recommendation: match → exposure → selection, with a separate path for brands that appear without observable live evidence. It is a measurement framework, not a claim about a secret AI algorithm.
Can AI find evidence for this buyer's specific question?
The onsite signal was far too large to ignore
Our large panel contained 34,960 unbranded GPT and Gemini monitoring observations from 75 anonymized projects. These were monitoring runs, not 34,960 independent conversations with buyers. We checked whether the target brand appeared in the answer, whether the source list cited its own domain and whether the AI-generated search queries named that brand.
On GPT, when the brand's own site was cited but the search query did not name the brand, the brand appeared in 49.0% of answers. When neither signal occurred, it appeared in just 2.8%. On Gemini the corresponding rates were 58.4% and 3.8%.
When GPT cited the brand's own site, the brand appeared far more often
Unbranded prompts; neither comparison group contained a branded search fan-out.
We also looked at repeated runs of the same brand, prompt and engine where own-domain exposure changed. Among runs with no branded fan-out, having the site in the observed source list was associated with an average 40.2-percentage-point higher mention rate on GPT and 49.0 points on Gemini. These comparisons reduce differences in brand identity and prompt wording. They do not remove every confounder, and a citation may be recorded alongside the same answer-generation process as the mention.
There is a second nuance that makes the argument more balanced. In a separate 199-prompt page-corpus analysis, matching the request to a page on the brand's website was only weakly predictive of GPT exposure (AUC 0.545) and more predictive for Gemini (0.641). Just writing an excellent page does not guarantee it will be retrieved. Offsite signals, indexing and distribution may help get useful evidence in front of an engine. Neither paper establishes how much lift a specific new link or page will cause.
Why a single "#1 ranking factor" is too simple
Our fifth paper sent the same prompts to four live AI engines and found mean citation-URL overlap of just 0.8%, with approximately 85% of engine pairs sharing no cited URL. A useful page may appear in one engine's evidence and never enter another's observed source list. Those differences are invisible in a single, broad industry-recommendation score. Read the comparison.
Even a brand that gets named may not have won a buyer. Our fourth paper on purchase advice found recommendations in 77.6% of purchase-directed conversation episodes, but observable same-mission buyer follow-up in only 26.9%. Ranking, citation, recommendation and purchase are not interchangeable outcomes.
None of this contradicts Wills' basic insight that relevant coverage matters. It changes what the insight can reasonably tell us to do.
What I would do with both studies
Start with a real buyer question. For the fictional hotel, do not begin with "we need 50 travel backlinks." Begin with "quiet family hotel near Gare de Lyon with late check-in" and the other constraints that appear across real conversations.
Check your own evidence. Does the hotel's site actually explain room types, location and check-in policy? Can someone verify it quickly? If not, improve that page. This is onsite work.
Check the outside footprint. Which travel guides, review sites, directories and relevant publications show up in the searches for that need? Are they accurate, current and mentioning the brand? This is where Wills' offsite approach is genuinely useful.
Finally, test retrieval and the answer. Run the request across engines, observe available fan-outs and citations, and distinguish an exposure problem from a selection problem. If the content is relevant but never appears, producing yet another onsite article may not be the highest-value next step. If the engine cites the page but omits the brand, investigate the evidence and competing options.
For marketers, I would frame the budget choice as content when match is missing, distribution when exposure is missing, and testing when you do not know which stage failed. The offsite and onsite work are not rival religions.
One experiment that could resolve the debate
Choose brands and prompts with comparable baseline visibility. Measure the full search-and-answer path repeatedly. Improve onsite coverage for one randomized group, earn relevant offsite mentions for another, combine the two for a third, and retain a holdout. Then compare changes in retrieval, citations and brand mentions across GPT and Gemini. Control for time and report uncertainty. We would learn far more than from another ranking-factor correlation chart.
I would be happy to compare methods with Ben Wills. The most useful version of the debate is a shared measurement protocol that can tell marketers which intervention actually moved the outcome.
Sources and methods
- Ben Wills / OppAlerts, We Now Know the #1 Ranking Factor for AI and LLMs (12 October 2026). The published article describes the method and charts but does not, on its own, establish causal uplift.
- Tannenbaum, From Prompt to Recommendation (September 2026 preprint). See Sections 4–7, especially Table 3 and within-prompt comparisons.
- Answer-Reconstruction Search Density: an AI answer can compress multiple traditional search tasks.
- The Prompt Is Not the Query and Beyond the Final Prompt: the active buyer request may not fit in a single prompt.
- Purchase Advice and Observable Buyer Responses: recommendation is not a purchase.
- Scoring With the Engine: sources differ sharply across engines.