The Outrigger AI Visibility Index: What Study 1 Said, and What the v3 Paper Retains

This article was first published on 9 April 2026 as the Phase 1 report of the Outrigger AI Visibility Index, covering a Phase 1 sub-sample of 1,004 businesses. On 7 May 2026 the Phase 1 dataset record (DOI 10.5281/zenodo.20048614) and the v2 paper (DOI 10.5281/zenodo.20048424) were withdrawn, and the v3 paper, Off-Page Signals Have Model-Specific Effects on Generative AI Search Visibility, superseded them.

The rule we apply to our own research is: republish exactly or remove. The Phase 1 tables, thresholds and percentages that used to fill this article are not reprinted in the v3 paper, so they have been removed from this page. What remains is (1) what the v3 paper says about Phase 1, quoted with section numbers, (2) the v3 key findings, and (3) the quotations from Joel House that appeared in the original article, kept verbatim for the record and labelled as commentary on the superseded Phase 1 analysis.

The paper on Zenodo is the authoritative source.

> “This paper supersedes earlier reports of the same dataset. Version 1 (April 2026) reported headline statistics on a Phase 1 sub-sample of 1,004 businesses.” (v3 §1 Introduction)

> “Sampling proceeded in two waves. Phase 1 (n = 1,004) sampled businesses URL-first, beginning with an industry-city directory query and verifying first-party domain ownership. Phase 2 (n = 1,725) sampled name-first, beginning with a business name and resolving the URL post-hoc via Apify Google SERP.” (v3 §3.2 Sample)

The v3 paper prints exactly two figures from Phase 1: the sub-sample size and Domain Authority’s raw correlation on it.

> “Domain Authority has a raw Pearson correlation of 0.337 with the composite visibility score on the Phase 1 sub-sample (n = 1,004) and 0.217 on the full sample (n = 2,200 businesses with non-null DA). After BH-FDR correction it remains significant as a raw correlation. In the multivariate L2-regularised logistic regression on n = 2,729 — which simultaneously estimates 22 features — its standardised coefficient is +0.139 with a bootstrap 95% confidence interval of [−0.031, +0.316], which crosses zero.” (v3 §4.4)

Every other Phase 1 number from the April article (the invisibility percentage, the review thresholds, the Domain Authority bands, the per-model mention rates, the industry table, the llms.txt analysis, the sentiment split, and the self-assessment and tiered action plan built on them) traces only to the withdrawn Phase 1 record and is not republished.

Roughly seven in ten businesses (69.5% of 2,729) were not mentioned by any AI system they were tested on. Gemini and Google AI Overviews were tested for a subset of the sample only.

Off-page presence (directories such as BBB, Yelp and Trustpilot, Wikipedia, LinkedIn company pages, and Reddit and Quora mentions) correlates with being mentioned by ChatGPT, Claude and Perplexity after nine controls (v3 Tables 2–4).

Businesses cited as a source by Perplexity were roughly six times as likely to be mentioned in that response.

The sample was “2,729 businesses across 14 verticals” and “four metropolitan markets (Los Angeles, New York, Chicago, Sydney)” (v3 §3.2).

The paper’s Table 2 (Reddit mention count, partial correlation by model, nine controls) prints: Claude +0.190 [+0.147, +0.233], ChatGPT +0.128 [+0.081, +0.174], Perplexity +0.110 [+0.063, +0.156]. Tables 3 and 4 (Quora, Wikipedia) are reproduced on the Study 2 page. Rows for ChatGPT, Claude and Perplexity are shown.

> “The single largest off-page partial r anywhere in the table is has_wikipedia → Claude (+0.219), and the second largest is quora_mention_count → Claude (+0.203).” (v3 §4.2)

Businesses cited as a source by Perplexity were roughly six times as likely to be mentioned in that response.

> “Perplexity and Google AI Overview expose the URLs they cite as sources in their structured responses. ChatGPT, Claude, and Gemini do not (their sources_cited arrays are empty by API construction, not by absence of citations).” (v3 §4.3)

> “All of these findings are correlational. The dataset cannot adjudicate causality. A business that is mentioned on Reddit and on Wikipedia and that ranks in BBB and Yelp directories is likely to be a more established business with more brand equity, and brand equity itself plausibly drives both off-page presence and AI mention probability.” (v3 §5 Discussion)

> “Observational, not causal. The dataset cannot distinguish causation from confounding. Partial-correlation and OLS-residualisation analyses isolate linear contributions of a predictor controlling for a measured set of confounders; unobserved confounders (brand age, marketing spend, executive PR effort, demand-side prominence) are not controlled for.” (v3 §6 Limitations)

> “Single point in time. All probing was conducted in April–May 2026. AI model versions, training cutoffs, and grounding behaviour change frequently. Replication on the same sample at a later date would likely produce a different mention pattern at the margin.” (v3 §6 Limitations)

> “The author is the founder and operator of Outrigger, a generative engine optimization platform whose product thesis overlaps with the research questions of this paper.” (v3 §3.1 Conflict of interest disclosure). Outrigger is now Outrigger.

The controlled trial the paper proposed to test causal magnitude (Phase 3, Layer 3) was pre-registered on OSF (ZFA38, 7 May 2026); the trial did not launch; no businesses were enrolled.

The quotations below appeared in the April 2026 article. They were made about the Phase 1 analysis, which the v3 paper has since superseded. Figures inside them refer to the withdrawn Phase 1 record; they are kept here verbatim as a record of what was said at the time, not as current findings. Where the v3 paper reached a different conclusion (for example on Domain Authority, see §4.4 above), the paper governs.

"Every SEO tool is trying to sell you a button to push," says House. "The data says there is no button. The two things that predict AI visibility are the two things that take the longest to build — domain authority and genuine customer volume. Nothing beats them, and nothing replaces them."

"If you can only invest in one thing, build domain authority," says House. "A business with DA 50 and 30 reviews is more visible to AI than one with DA 15 and 3,000 reviews. That’s the opposite of what most ‘AI SEO’ advice will tell you, but it’s what the data shows every time we run the numbers."

"The jump at 1,000 reviews isn’t magic," says House. "It’s the point where your review count stops looking like ‘this business runs a review campaign’ and starts looking like ‘this business has real customer volume.’ That’s the signal the models are picking up on."

"You can’t optimize for ‘AI’ as one thing," says House. "ChatGPT, Perplexity, Gemini, and Claude disagree with each other 89% of the time on who to recommend. They’re not a single recommendation layer — they’re five different audiences with different biases, and you need a different strategy for each."

"Every SEO tool in the market right now is selling llms.txt as the AI visibility silver bullet," says House. "We tested it against 1,004 businesses. The honest answer is it only helps if you’re in the middle — DA 21 to 60. If you’re small, it doesn’t do anything. If you’re already big, you don’t need it. The industry is selling a mid-market play as a universal fix."

"Here’s a reframe most marketers haven’t processed yet: there is no such thing as bad AI press," says House. "The models don’t write negative reviews. They either recommend you or they pretend you don’t exist. The entire risk equation shifts from reputation management to presence management."

"If you run a med spa, a dental practice, or a home services business, your industry is structurally disadvantaged for AI visibility," says House. "The entire category has low domain authority and local-only presence. The good news is that means your competition is also invisible — the first mover in each local market gets a near-monopoly."

"The businesses winning at AI visibility aren’t doing anything clever," says House. "They’re not hacking prompts. They’re not gaming schema. They’re the same businesses that built real authority the old-fashioned way. AI models are rewarding the same thing Google has always rewarded — they’re just rewarding it louder."

"The reason most agencies can’t help clients with AI visibility is that they can’t measure it," says House. "You can’t improve what you don’t track. Run the 60-second test every month, across four models, with the same prompts. Watch the number move. That’s the only honest feedback loop in this space."

Cite the paper, not this article: *House, J. (2026). Off-Page Signals Have Model-Specific Effects on Generative AI Search Visibility: Evidence from a Cross-Platform Audit of 2,729 Businesses Across Five Generative AI Systems (v3). Zenodo. https://doi.org/10.5281/zenodo.20076380*.

Key takeaways and methodology are published on the Outrigger AI Visibility Index research hub. If you have questions about the methodology or findings, reach us through our research contact form. For press, podcast appearances, or speaking engagements related to this research, get in touch through our contact form.

Read the paper. Then run the simplest version of the test on your own brand: ask ChatGPT, Perplexity, Gemini and Claude the buying-intent questions your customers ask, and record whether you appear.

If you want the measured version, run an AI visibility audit: it establishes your baseline across the four AI assistants and gives you a prioritised plan. That baseline is the first move for any brand that is currently absent from the answer.