What off-site signals matter for AEO?

Off-site signals carry weight because answer engines assemble most of their responses from pages nobody at your company controls. Muck Rack's July 2025 analysis of millions of AI-cited links across ChatGPT, Gemini, and Claude found that more than 95% of citations came from unpaid sources, with 85% of those classified as earned media [3]. Peec AI's March 2026 study of 30 million sources across five AI systems ranked Reddit, YouTube, and LinkedIn as the most-cited domains overall, with Yelp and G2 appearing heavily on recommendation queries [7]. Your own domain is one voice in that mix, and in most categories a minority one.

What the evidence base actually looks like

Almost everything published about off-site signals is correlational rather than causal. Ahrefs ran Spearman correlations across 75,000 brands in December 2025 and found branded web mentions at 0.664 against 0.266 for Domain Rating [2], then flagged the limitation themselves: "correlation isn't causation. We've spotted patterns between search metrics and AI mentions, but that doesn't mean improving these metrics will automatically boost your AI visibility" [2]. Large brands get mentioned more or less everywhere, so part of what a strong correlation measures is simply size.

Platforms document eligibility, not selection

Google's documentation covers eligibility and stops short of selection. To appear as a supporting link in AI Overviews or AI Mode, "a page must be indexed and eligible to be shown in Google Search with a snippet," and Google adds that "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary" [1]. Which of the eligible pages gets pulled into a given answer is inferred, not published.

Branded mentions are the most consistently reported correlate of AI visibility, and they outscore link metrics in every dataset that measures both. Ahrefs put branded web mentions at 0.664 against 0.266 for Domain Rating and roughly 0.2 for raw backlink counts, with YouTube mentions the strongest single signal at 0.737 [2]. Semrush's study of 1,094 US categories tracked in ChatGPT between January and June 2026, covering more than 50,000 brands, found branded search volume was the only traditional metric to reach statistical significance as a differentiator between category owners and everyone else, and even then the edge was modest at 55.7% of comparisons [9].

Read together, those findings point at something narrower than "get mentioned more often." Mentions appear to work as evidence that other people are discussing an entity, which fits branded search volume and YouTube mentions both tracking closely. The practical version is to get your brand name into places where humans already talk about your category: podcasts and video with genuine audiences, industry newsletters, conference recaps, customer stories published on someone else's site. Piling up mentions on low-traffic pages nobody reads has no observed relationship with anything.

On measurement, track mention volume and AI mention rate as separate series across at least two quarters, then check whether they move together for your brand specifically rather than across the market.

Reviews are the off-site signal with the clearest buyer-side evidence. G2's March 2026 survey of 1,076 B2B decision-makers found 45% named a review-site citation the most confidence-inspiring signal inside an AI answer, rising to 50% among daily AI chatbot users [4]. G2's July 2026 research found review sites had become the top source shaping which vendors make a shortlist at 38%, edging past AI chatbots at 37% for the first time [5]. BrightLocal's 2026 survey of 1,002 US adults found 97% read reviews and 47% will not use a business with fewer than 20 reviews [6].

Recency is the part most teams get wrong

BrightLocal found 74% of consumers look for reviews written in the last three months, and 32% want reviews from the last two weeks, up from 20% a year earlier [6]. A five-year-old pile of 200 reviews reads differently to both a person and a summarizer than 30 reviews from this quarter, which makes review velocity a separate metric from review count.

What that looks like in practice

Attach the ask to a process that already runs, such as onboarding milestones or renewals, rather than treating reviews as a quarterly campaign. Concentrate on the two or three platforms your category actually gets cited from. Respond publicly to negative reviews, since 82% of consumers now read AI-generated review summaries that compress sentiment across the whole corpus into a sentence or two [6].

Comparison pages and category rankings are the format AI systems reach for on recommendation prompts, because those pages already contain the structured judgment the model is being asked to reproduce. Peec AI's 30-million-source analysis found G2 and Yelp appearing heavily on recommendation-type queries, with Perplexity leaning on G2 for B2B prompts specifically [7]. A June 2026 synthesis by 5W, a public relations firm, combined nine citation datasets and found extreme concentration inside verticals, with TechRadar alone accounting for 8.86% of B2B SaaS CRM citations [10].

That concentration is what makes this signal workable. In most categories a small set of pages supplies a large share of the recommendation-shaped answers, so finding them is a finite research task rather than an open-ended content program. Run your core buying prompts, log every URL that gets cited, and you will usually end up with five to fifteen pages doing most of the work.

What you can do about each page varies. Some are editable through a vendor profile you already control, where category placement and feature data are yours to fix. Others are journalist-written roundups where the only route in is a pitch carrying something worth writing about. A third group are affiliate listicles with commercial terms attached, and those carry a caveat: paid and advertorial content made up a negligible share of AI citations [3].

Community platforms are the most-cited source category in nearly every multi-platform dataset, and also the least stable. Peec AI ranked Reddit the single most-cited domain across five AI systems [7], and the 5W synthesis put it at an estimated 46.7% of Perplexity's top-10 source share [10]. Semrush's thirteen-week study of 230,000 prompts watched Reddit's presence in ChatGPT responses fall from roughly 60% in early August 2025 to about 10% by mid-September while holding steady elsewhere [8].

This area also contains one of the few causal tests on record. Andrew Shotland of Local SEO Guide ran a month-long campaign using 1,000 purchased Reddit accounts to place 100 brand mentions and 100 comments in relevant threads for a SaaS client, and citation rates in Google AI Overviews rose from roughly 8 to 9% up to roughly 24 to 27% across 80 tracked prompts [14]. The more useful finding is what happened afterward: visibility "went straight back down" once the campaign stopped [14].

Community discussion clearly influences answers, and manufactured community discussion influences them temporarily while creating exposure to platform enforcement. The version that holds up is slower: employees participating under real identities in the forums where your category gets discussed, answering questions without a pitch attached, and saying something specific enough that a model has a reason to quote it.

Earned coverage supplies the largest measured share of what AI systems cite. Muck Rack's July 2025 study, built from a large prompt set run across web-enabled versions of ChatGPT, Gemini, and Claude, found more than 95% of citations came from unpaid sources, 85% of those were earned media, 27% were journalistic, and half of all AI responses included at least one earned media citation [3].

Which outlets earn those citations is where expectations tend to break. The 5W synthesis of nine datasets found that the Wall Street Journal, the New York Times, and Bloomberg did not crack the top twenty most-cited domains, while trade and vertical publications dominated inside their own categories [10]. A quote in a national business paper is a real achievement with limited AI citation value, whereas placement in the trade publication your buyers already read is less impressive and considerably more useful.

The work is to build a short target list of publications that already appear in your own citation logs, then supply what they actually run: original data from your operations, a practitioner position on something contested in your field, or a named expert available for comment on a beat. Earned media is the slowest of the six signals, usually two to four quarters before movement shows up in citation tracking.

Entity consistency is the only signal on this list where platform documentation is explicit, and also the one with no published effect size. Google's Organization structured data documentation describes sameAs as "The URL of a page on another website with additional information about your organization," offering "a URL to your organization's profile page on a social media or review site" as the example, and notes that some properties are "used behind the scenes to disambiguate your organization from other organizations" [11]. Google's business profile guidelines separately require that a business name "reflect your business's real-world name, as used consistently on your storefront, website, stationery, and as known to customers" [12].

No study has shown that fixing entity data raises AI citation rates by a measurable amount. The argument for doing it anyway is cost. Reconciling your legal name, trading name, and common abbreviations across your site, your schema, and your review profiles is a few days of work with a permanent shelf life, and it removes a failure mode where two versions of the same company get treated as unrelated entities that each accumulate half the evidence.

In practice, settle on one canonical name and description, then make every profile match it. An Organization block with sameAs pointing at the profiles you genuinely maintain gives engines a machine-readable map between them, and an internal rule about permitted name variants stops the problem returning.

Four kinds of evidence exist in this area, and they support very different levels of confidence.

Documented platform behavior

The strongest tier and by far the smallest. Google states the indexing and snippet eligibility requirement for AI Overviews and AI Mode plainly [1] and documents its Organization and business profile fields [11][12]. Documentation of this kind tells you what is required; it almost never tells you what is weighted.

Controlled manipulation

Rare, and more revealing than its sample size suggests. The Reddit campaign above is close to a genuine causal test, with a treatment, a fixed prompt set, and a measured reversal once treatment stopped [14]. It covered one client and 80 prompts, so it establishes that the channel works without indicating what it is worth in your category.

Large-sample correlation

The bulk of what circulates in this field. Ahrefs' 75,000-brand study [2] and Semrush's 1,094-category study [9] are large enough to be meaningful, and both come from companies that sell AI visibility tracking. Neither controls for brand size, the obvious confound behind every mention-related coefficient.

Buyer surveys

Useful for a different question. The G2 [4][5] and BrightLocal [6] surveys measure what buyers say influences them, which predicts behavior imperfectly but does describe the environment your citations land in.

One caution runs across all four tiers: citation share is volatile and platform-specific. Semrush watched Wikipedia fall from roughly 55% of ChatGPT responses to under 20% in six weeks while staying flat elsewhere [8], and only around 11% of domains are cited by both ChatGPT and Perplexity [10].

Order the six signals by evidence-to-effort ratio rather than correlation strength, because the strongest correlate is also the slowest and most expensive to move.

Start with entity consistency

The cheapest item on the list, resting on documentation rather than inference, and it makes every later signal accumulate against a single entity instead of two partial ones [11][12]. Budget a few days of reconciliation, then a maintenance rule.

Then the review corpus

For B2B and local businesses this carries the best evidence-to-effort ratio, with survey data from both sides of the transaction [4][5][6]. Keep the flow steady rather than lumpy, given that 74% of consumers want reviews from the last three months [6].

Then the pages already being cited in your category

A finite research task. Run your buying prompts, list the URLs that come back, and sort them by whether you can edit them, pitch them, or neither. Concentration inside verticals is high enough that the list is usually short [10].

Then earned media and mentions as a standing program

The strongest correlate in the data, and also the slowest to move [2][3]. Commit to two to four quarters or leave it alone, since a one-quarter push produces little you can measure.

Narrow the target before you spend

Semrush found that only 15.2% of 1,094 tracked categories had a clear brand owner, and that once a brand held a category it kept the lead in 90.4% of month-over-month comparisons [9]. The highest-volume topics were the most fragmented, with just 11.3% showing a clear owner [9]. Picking a narrower category you can plausibly own tends to beat spreading the same budget across a broad one.

Most of the correlation data comes from vendors

Ahrefs, Semrush, and the AI visibility platforms publishing citation studies all sell tools that measure the thing their research says matters [2][8][9]. Sample sizes are real and methodology is usually disclosed, but no independent replication of the headline correlations exists yet.

Manufactured signals decay

The one reversal test on record showed AI Overview citation gains disappearing as soon as the campaign stopped [14]. Anything purchasable in a month is equally purchasable by a competitor, and the platforms hosting the discussion run their own enforcement against it.

AI systems misattribute what they cite

The Tow Center at Columbia Journalism Review asked ChatGPT to identify the source of 200 quotes from 20 publishers and received partially or entirely incorrect responses on 153 occasions [15]. Off-site signals can be working while the resulting citation lands on the wrong page or credits the wrong company.

Review volume cuts both ways

Volume does nothing if the ratings are weak. BrightLocal found 31% of consumers will only use a business rated 4.5 stars or higher [6], so a review push layered on unresolved service problems accelerates an outcome you did not want.

Wikipedia is not a lever you can pull

Wikipedia's notability guideline states that "Self-promotion, autobiography, product placement, press releases, branding campaigns, advertisements, and paid material are not valid routes to an encyclopedia article" [13]. A Wikipedia entry follows independent coverage rather than producing it.

Not link building under a new name

Link acquisition and mention acquisition overlap, but they score differently in the data. In the largest correlation study available, raw backlink counts sat around 0.2 while branded mentions reached 0.664 [2]. A mention with no link attached still appears to count.

Not paid placement

More than 95% of AI citations came from unpaid sources in Muck Rack's analysis, with paid and advertorial content making up a negligible share [3]. Sponsorship buys attention from people, not citations from engines.

Not a substitute for being findable

Off-site investment aimed at a site that is blocked, broken, or unindexed changes nothing, because Google requires a page to be indexed and snippet-eligible before it can support an AI answer [1].

Not a fixed ranking formula

These are observed patterns across platforms that behave differently and shift month to month [8][10]. Anything presented as a weighted off-site scoring model for AI visibility is an inference drawn from correlation, not a published algorithm.