SEO is becoming brand engineering, not ranking engineering

Executive summary

For twenty years, search engine optimization was an engineering discipline pointed at a single output: position. You built links, you tuned pages, you climbed. In 2025 and 2026, five separate, methodologically independent studies — from Ahrefs, Semrush, SparkToro, Omniscient Digital, and Profound — quietly converged on a conclusion that breaks that model. The strongest correlate of whether an AI engine names your brand is not your backlink profile and not your ranking position. It is how much, and where, you are talked about off your own website.

Ahrefs studied 75,000 brands and found branded web mentions correlated with AI visibility at roughly 0.66–0.71, while backlinks limped in at 0.17–0.29. (Confidence: High.) Omniscient Digital read 23,387 AI citations and found 48% came from earned media and only 23% from brands' own sites. SparkToro and Gumshoe.ai ran 2,961 prompts and found roughly 85% of brand mentions came from third-party pages. Semrush and Kevin Indig found that 62% of AI citations don't even name the brand they're built from.

The uncomfortable part isn't that any one of these is true. It's that all of them are true at once, they were produced by people who don't coordinate, and they describe a search economy that the field's own vocabulary — "ranking factors," "link equity," "position one" — can no longer explain. Meanwhile Google keeps telling the industry that "normal SEO" is all anyone needs. This piece is about the gap between those two stories, and why closing it means treating your brand as something you engineer rather than something you earn by ranking.

A timeline of the inversion

The number that should have ended an argument

Start with the Ahrefs study, because it is the largest and the cleanest. Across 75,000 brands, researchers Louise Linehan and Xibeijia Guan measured the Spearman correlation between AI visibility and a battery of familiar SEO signals. Branded web mentions landed at 0.656–0.709. YouTube mentions were the single strongest factor at roughly 0.737. Backlinks — the currency the entire link-building industry is denominated in — came in at 0.174–0.289. Their own one-line summary was blunt: "Link and content volume has little impact on AI visibility." (Confidence: High — primary source, large sample, stated methodology.)

Correlation is not causation, and Ahrefs said so. But here is what makes the finding hard to wave away: four other teams, using completely different methods, landed in the same place.

• Omniscient Digital didn't measure correlation at all. It read citations — 23,387 of them across 240 branded prompts and five engines — and counted where the cited material physically lived. Result: 48% earned media, 30% commercial brand content, just 23% owned. On customer-review queries, earned media hit 82%. "When someone asks an LLM about your brand," said analyst Cate Dombrowski, "most of what it references comes from outside your website." (Confidence: High.)
• SparkToro and Gumshoe.ai went stranger. Across 2,961 runs they found AI brand rankings are effectively random — Rand Fishkin's framing: "there's a <1 in 100 chance that ChatGPT or Google's AI, if asked 100X, will give you the same list of brands in any two responses." And yet a stable consideration set of brands surfaced in 55–77% of responses, with about 85% of those mentions sourced from third-party pages. (Confidence: High.)
• Semrush + Kevin Indig measured the gap between being cited and being named. Across 3,981 domain appearances, 62% were "ghost citations" — the source was linked, but the brand was never mentioned. Citation rate 74.9%; mention rate 38.3%; both at once only 13.2%. (Confidence: High.)

Four research teams. Four methods — correlation, citation-counting, randomization testing, citation-vs-mention analysis. One conclusion: the thing that moves AI visibility is durable, distributed, off-site presence. Not the link graph. Not position.

Screenshot to include: A side-by-side bar chart titled "Correlation with AI visibility (Ahrefs, 75k brands)" showing branded web mentions (~0.68) and YouTube mentions (~0.74) towering over backlinks (~0.23). Caption it with the one-line Ahrefs verdict.

Where Google's story breaks

Now hold that data against what Google has been saying on stage for the entire same period.

At the Search Central Live APAC Deep Dive in Bangkok in July 2025, Gary Illyes told a near-capacity room that AI Mode and AI Overviews run on "the same crawling, indexing and serving infrastructure as traditional Search" and that "there's not much need for change since the mechanisms are largely the same as traditional search." (Confidence: High.) In December 2025 in Zurich, John Mueller offered the now-canonical line: "Good GEO is just good SEO." (Confidence: High.) In May 2026, Google formalized the position in documentation, writing that "optimizing for generative AI search is optimizing for the search experience, and thus still SEO." (Confidence: High — quoted via Search Engine Journal's coverage of the guide; re-confirm against the primary Search Central doc before publication.)

Read literally, "it's still SEO" is defensible. Crawling and indexing genuinely do still matter; nobody disputes that an uncrawlable page can't be cited. But notice what the framing quietly does: it pushes the conversation toward the part of the stack Google controls and can advise on — your own pages — and away from the part the data says actually decides AI visibility, which is what other people's pages say about you.

There's a sharper inconsistency underneath. Google's John Mueller has stated for years that Google doesn't use unlinked "brand mentions" as a ranking signal. (Confidence: Medium — this is paraphrased from longstanding secondary coverage; the exact original quote should be re-verified.) Yet the 2025–2026 correlation studies find branded mentions are the single strongest correlate of AI visibility. Both can be technically true if you're careful: "brand mentions aren't a classic ranking signal" describes the ten blue links; "brand mentions correlate with AI citation" describes the answer layer. But that's exactly the point. The truth has bifurcated by surface. Telling practitioners "it's all still SEO" papers over the fact that the winning move on one surface is nearly irrelevant on the other.

Counter-argument, taken seriously: Google is not lying, and the skeptics have a real case. Correlation studies can't separate "mentions cause citations" from "good brands attract both mentions and citations." Search Engine Land itself notes backlinks "remain one of the most influential ranking factors in 2026" for the classic SERP. (Confidence: Medium.) And AI visibility tools have an obvious commercial incentive to declare backlinks dead and a new metric — theirs — essential. A disciplined reader should treat every vendor study, including the ones above, as a hypothesis with a sales motive attached. The reason to believe it anyway is convergence: the vendors compete with each other, used different data, and still agree.

The dot nobody connects: citation ≠ recommendation ≠ mention

Here's the synthesis the individual studies don't state, because each only owns one piece of it.

The industry has been selling "AI visibility" as a single thing. The data says it's at least three things that move independently:

1. Citation — your URL appears as a source. (Semrush: happens 74.9% of the time for appearing domains.)
2. Mention — your brand name is spoken in the answer. (Semrush: only 38.3%.)
3. Recommendation — the model actually tells the user to pick you.

These come apart in counterintuitive ways. Kevin Indig's finding is that "strong brands get named without citation; aggregators get cited without naming." In his words: "The AI knows the information came from somewhere, but doesn't feel the need to explicitly say so. The brand name carries on its own." (Confidence: High.) A Wikipedia-grade entity gets recommended on the strength of what the model already absorbed about it in training — no live citation required. A thin affiliate listicle gets cited as a convenient source and gives all its visibility away to the brands listed inside it.

That reframes the entire objective. If you optimize for citations — the thing the GEO tooling industry sells dashboards for — you can "win" on a chart while losing in the answer, because 62% of those citations never say your name. The asset that actually compounds is the one that lets a model name you without needing a source: a strong, well-defined entity, reinforced across enough independent third-party surfaces that the model has internalized "this brand belongs in this category."

That is not ranking. That is brand engineering.

"Brand engineering" — a working definition

If ranking engineering optimized a page's relationship to a query, brand engineering optimizes an entity's relationship to a category in the model's representation of the world. Concretely, it means deliberately constructing four things:

• Entity clarity. A machine-resolvable identity — consistent name, category, founders, products, and a clean knowledge-graph footprint — so the model knows what you are and what set you belong to. (This is the one part that overlaps with classic technical SEO: structured data, an unambiguous "about," sameAs links.)
• Distributed corroboration. The same true claims about you, repeated across many independent third-party surfaces — earned media, expert roundups, podcasts, YouTube, community threads. Omniscient's 48% earned-media share and SparkToro's 85% third-party share are the targets you're feeding.
• Category association. Being named alongside the right competitors and in the right context often enough that the model files you into the consideration set. SparkToro's 55–77% consideration-set stability is what you're trying to enter.
• Recall over rank. Optimizing for the model to remember you, not for a position that, per SparkToro, won't even repeat twice.

Diagram to include: A funnel inverted into a network. Old model (linear): Keyword → Page → Link → Rank → Click. New model (web): a central brand entity node surrounded by earned-media, YouTube, Reddit, and Wikipedia nodes, each feeding "corroboration" edges inward, with three differently-weighted output arrows labelled Cited, Mentioned, Recommended.

The Reddit problem: your best channel is one you don't own and can't trust

The brand-engineering thesis has an ugly dependency. The single richest vein of off-site brand-talk — community discussion, with Reddit at its center — is simultaneously the most influential, the most contested, and the most manipulable.

Reddit became the most-cited source in Perplexity's answers, then sued Perplexity in October 2025 over how that content was being scraped; Profound later measured Perplexity's Reddit citations dropping roughly 86% in the aftermath. (Confidence: Medium — the 86% figure is reported via secondary summaries of Profound's data; confirm on Profound's dashboard before quoting.) So the channel that most moves your AI visibility can be reshaped overnight by a lawsuit between two companies you have no relationship with.

It gets worse, and this is the bridge to the harder story: because brand-talk on UGC platforms is what the models read, seeding that talk is now a documented attack surface — one researchers have shown can be poisoned with as few as 13 words. The same mechanism that makes brand engineering work makes it abusable. (We treat that in depth in the companion piece on recommendation poisoning.) For now, the strategic point is narrower: if your off-site presence is your most important asset and you don't manage it, someone else — a competitor, an affiliate, or a bad actor — will manage it for you.

What everyone is missing

Three things, stacked.

First, the industry is measuring the wrong object. AI "rank trackers" sell a number that SparkToro showed is statistically random — Fishkin's verdict on rank tracking in this context was simply "full baloney." (Confidence: High.) Chasing a position that won't repeat twice is worse than useless; it's a distraction dressed as rigor. The measurable object that survives is share of consideration set — how often you appear at all, across many runs — which is exactly what Semrush's 126-million-prompt AI Visibility Index is now built to track.

Second, "it's still SEO" is true in a way that makes it misleading. Crawlability still matters, so the statement passes a fact-check. But it routes attention to the 23% of citation weight that lives on your own site and away from the 77% that doesn't. The honest version is: the foundations are still SEO; the competitive advantage has moved off-site.

Third — and this is the part nobody states plainly — the new game has lower marginal returns to effort and higher returns to identity. You can quadruple your content output and barely move AI visibility (Ahrefs: content volume has little impact). But becoming the brand that's associated with a category compounds: once a model has internalized you, it names you without a citation, on queries you never targeted, in a consideration set you didn't bid on. That's a different kind of moat than the link graph ever was — slower to build, far harder for a competitor to copy with a budget.

Future predictions

• Rank trackers rebrand or die. By late 2026, "AI rank tracking" gives way to share-of-model / consideration-set metrics as vendors quietly concede ordering is noise. (Confidence: Medium.)
• PR and SEO budgets merge. The function that produces distributed third-party corroboration — historically PR and "digital PR" — absorbs a growing share of what used to be link-building spend. (Confidence: Medium.)
• Entity infrastructure becomes table stakes. Knowledge-graph hygiene, Wikipedia/Wikidata presence, and consistent entity markup stop being "advanced SEO" and become the baseline a brand needs just to be resolvable. (Confidence: High.)
• Google keeps saying "it's still SEO" — and keeps shipping evidence it isn't. The messaging holds because it's defensible; the product roadmap keeps undercutting it. (Confidence: Medium.)

Practical takeaways

1. Audit your entity before your pages. Can a model state what you are, what category you compete in, and who your peers are? If not, fix entity clarity (structured data, a clean "about," Wikidata/Wikipedia footprint) first.
2. Move budget from link volume to mention distribution. The Ahrefs correlation gap (0.68 vs 0.23) is your reallocation thesis. Fund earned media, expert inclusion, podcasts, and especially video.
3. Treat YouTube as a primary AI-visibility channel, not social media. It was the single strongest correlate in the Ahrefs data and the most-cited domain in AI Overviews. Transcripts are training-and-retrieval fuel.
4. *Engineer for mention, not just citation. Track whether your brand is named* in answers, not just whether a URL is linked. Remember 62% of citations never say the name.
5. Stop reporting AI "rank." Report share of consideration set. Run the same prompts many times; measure presence frequency, not position.
6. Manage your community footprint deliberately. Reddit and forum discussion is your highest-leverage and least-controlled asset. Monitor it, participate honestly, and treat misinformation about your brand as a visibility emergency.