There Is No 'AI Search' — Why AEO Practitioners Need to Stop Treating It As One

Lily Ray published a study this week tracking 11 websites that lost organic rankings after Google's January 2026 algorithm update. Every one of them also lost AI citations. Google AI Mode down 23.8%. ChatGPT down 27.8%. The cascade was consistent. Except on Perplexity — where the average decline was just 2.9%, and seven of the eleven sites actually grew.

That gap — 27.8% versus 2.9% — is not noise. It's a structural fault line. It means what we've been calling "AI search" is not one system. It's at least three, each operating on different logic, drawing from different sources, and responding to different signals.

This week, several independent researchers published data that — laid side by side — reveals this fault line from multiple angles. None of them framed it quite this way. Here is what we see when the pieces are assembled. We're calling this the two-reality framework — not because it's permanent, but because it's the clearest description of what the data currently shows.

Reality 1: Google's AI is still Google

Google AI — both AI Overviews and the newer AI Mode — is fundamentally downstream of Google's existing ranking system. If you lose organic, you lose AI.

Lily Ray published a study in February 2026 tracking 11 websites that lost organic visibility during Google's January 2026 algorithm update. Every single site also lost AI citations. Google AI Mode citations fell by an average of 23.8%. ChatGPT — which we'll come back to — fell by 27.8%. The cascade was consistent across all 11 sites, not a one-off.

Ray's conclusion: organic ranking degradation automatically cascades to AI citations.

Separately, Averi.ai's B2B SaaS citation benchmarks report found a 76.1% correlation between Google AI Overview citations and organic top 10 rankings (compiled in Onely's analysis). Semrush's AI Mode study found ~53% domain overlap and ~35% URL overlap between AI Mode citations and the organic top 10.

Those two numbers — 76.1% correlation and 53% overlap — seem to diverge, but they measure different things. Averi.ai measures statistical correlation across a large dataset; Semrush measures literal domain overlap in cited results. Both point the same direction: significant but not total dependence on organic rankings. The gap — the 47% of AI Mode domains that DON'T appear in the organic top 10 — is itself instructive. Google AI is drawing from organic but supplementing with its own sources. Semrush's AI Overview research found that Reddit, Quora, and YouTube are frequently cited before traditional websites.

The implication for practitioners: for Google's AI products, the primary optimisation strategy remains traditional SEO. The principle that AEO requires SEO fundamentals — a point Google's John Mueller has reinforced — now has quantitative evidence behind it, and a warning. Ray is explicit: don't invest in AEO tactics that could be detrimental to SEO performance. The cascade runs one way, and it starts with organic.

If the story ended here, the prescription would be simple: invest in SEO, and AI visibility will follow. But the story doesn't end here.

Reality 2: independent AI operates independently

Perplexity and — to a lesser but real extent — ChatGPT do not follow Google's rules.

Return to Ray's 11-site study. While Google AI Mode citations dropped 23.8% and ChatGPT dropped 27.8%, Perplexity showed only a 2.9% decline. Seven of the eleven sites actually grew their Perplexity citations while losing Google organic rankings. The sites that became less visible on Google became more visible on Perplexity.

This is not a marginal difference. It's a fundamentally different system.

Averi.ai's citation benchmarks make the divergence quantitative: only 11% of domains receive citations from both ChatGPT and Perplexity. The platforms favour different sources entirely. ChatGPT's most-cited domain is Wikipedia, accounting for 47.9% of its top citations. Perplexity's most-cited source category is Reddit, at 46.7%. These are not variations on a theme. These are different citation engines with different input preferences.

ChatGPT sits in between the two realities. Ray's data shows it's heavily dependent on Google's search results today — its 27.8% citation drop tracked closely with organic losses. But OpenAI is building its own search index. ChatGPT's citation behaviour will likely shift as that index matures. For now, optimising for Google mostly covers ChatGPT too. That may not hold.

The strategic gap is Perplexity. Optimising for Google — even perfectly — gives you Google AI visibility and probably ChatGPT visibility. It does not give you Perplexity visibility. Perplexity requires community presence: Reddit discussions, forum threads, UGC mentions — signals that have nothing to do with traditional SEO. For any business whose audience uses Perplexity (and its usage is growing), a Google-only strategy leaves a platform-sized blind spot.

What this means for measurement

If the platforms are independent, a blended "AI visibility score" is meaningless. Imagine telling a client they have "moderate AI visibility" — a 40% brand frequency on ChatGPT, 5% on Perplexity. On ChatGPT, they're one of the most-recommended businesses in their category. On Perplexity, they barely exist. "Moderate" hides both facts. The diagnosis is different. The fix is different. The blended number conceals the one thing the client actually needs to know.

Rand Fishkin's research with Near Media offers a path forward. Working with 600 volunteers across Claude, ChatGPT, and Google AI, Fishkin found that while individual AI responses are unpredictable, brand mention frequency becomes statistically meaningful when you run 65-90 prompts per platform. Traditional rank tracking — checking where you appear in a single AI response — is unreliable to the point of uselessness. Frequency tracking across dozens of prompts works.

Fishkin's data also reveals a market-size effect. In narrow local markets — his example was Los Angeles Volvo dealers, a set of roughly 9-11 businesses — the top four businesses clustered tightly and dominated recommendations. In broad categories like science fiction novels, he observed 211 unique titles across 99 responses. The smaller the competitive set, the more measurable the results.

This has direct implications for local service businesses. A furniture maker competing in a London neighbourhood is closer to the Volvo dealer scenario than the sci-fi novel scenario. The market is narrow. The competitors are countable. Visibility gains should be detectable at relatively low prompt counts.

The measurement framework that follows from this: per-platform, frequency-based, run at sufficient sample size. A business might show 40% brand frequency on ChatGPT and 5% on Perplexity. Those two numbers tell entirely different stories and demand entirely different interventions.

The quality signal is measurable

This isn't only about which platform you optimise for. It's about what kind of content survives across all of them.

Lily Ray's second study this week examined sites penalised for self-promotional listicles — content where companies ranked themselves first without independent evaluation, often generated entirely by AI. The visibility losses were severe: -42% to -49% across affected sites. These sites lost both organic rankings and AI citations simultaneously. Content that 100% triggered AI detection tools, programmatic templates scaled across hundreds of pages, and misused schema markup (fabricated AggregateRating, for instance) were common characteristics.

Wil Reynolds of Seer Interactive put it sharply in a February 2026 interview: "People are mortgaging their trust to get back a little extra time."

The trust penalty is no longer just a reader experience issue. It functions as an algorithmic signal — content that shortcuts quality loses the community engagement (shares, discussions, references) that platforms like Perplexity use for citation decisions. The readers Reynolds is worried about losing are the same readers whose behaviour feeds the citation algorithms.

On the other side of the quality spectrum, Digital Bloom's AI citation research found that content with attributed statistics receives a 22% AI visibility lift. Expert quotations deliver a 37% lift. These are signals of genuine expertise — original data, named authorities, verifiable claims. They're the structural opposite of AI-generated filler, and the citation data suggests that the difference is measurable.

A framework for what comes next

So what do you actually do on Monday morning? Five recommendations that follow from the evidence:

1. Stop selling "AI visibility" as one thing. Report per-platform. Diagnose per-platform. Recommend per-platform. A client's ChatGPT visibility and their Perplexity visibility are independent variables with different drivers.

2. Measure with statistical rigour. Run 65-90 prompts per platform and track brand mention frequency over time. A single AI query is an anecdote. Sixty-five queries per platform is data.

3. Invest in distributed authority. On-site SEO gets you Google AI and probably ChatGPT. Perplexity requires presence you don't control — reviews, community discussions, press mentions, forum threads. The Search Engine Land authority framework calls this "distributed authority," and it's the hardest signal for competitors to manufacture.

4. Don't skip SEO fundamentals. Lily Ray's cascade finding is unambiguous: all 11 sites that lost organic rankings lost AI citations too. SEO is not the old way. It's the foundation on which AI citation is built.

5. Treat quality as measurable infrastructure. Attributed statistics (+22% visibility lift), expert quotations (+37%), answer-first formatting, content freshness (65% of AI bot traffic targets content published within 12 months) — these aren't editorial preferences. They're quantitatively associated with higher AI citation rates.

The convergence objection

The strongest counter-argument isn't that this data is wrong — it's that the divergence is temporary. Web search in 2004 showed similar platform fragmentation across Google, Yahoo, and AltaVista. Different crawlers, different indexes, different ranking logic. Competitive pressure drove convergence within years. The same forces may be at work now: ChatGPT is already incorporating web freshness signals, Perplexity is already using traditional ranking heuristics alongside community signals, Google AI is already pulling from sources beyond its organic index. That 11% domain overlap could be 55% in eighteen months.

There's a sharper objection: even accepting the current divergence, the foundational advice converges anyway. Create authoritative, well-sourced content. Build genuine community presence. Maintain strong organic fundamentals. A practitioner who does those three things well performs across all platforms. The "three different systems" framing may create analytical complexity without changing what you actually do.

Both objections are reasonable. We think the convergence timeline is uncertain enough — and the current measurement gap damaging enough — that platform-specific reporting is worth the effort now, even if the platforms converge later. And while the foundational advice overlaps, the diagnosis of why a specific business is invisible on a specific platform requires platform-level data. "Do good SEO" is correct but insufficient when a client is visible on ChatGPT and invisible on Perplexity — the fix for the second problem is community presence, not more on-site optimisation.

---

Honest caveats

This analysis rests on a small number of studies published within a single week. Lily Ray's cascade study covers 11 websites — enough to establish a pattern, not enough to declare a universal law. The citation data compiled in Onely's analysis (sourced from Averi.ai and Digital Bloom) provides specific numbers, but from a limited number of research teams. Fishkin's volunteer study offers a measurement framework, but 600 volunteers is not a controlled experiment.

We're also looking at a snapshot. ChatGPT's citation behaviour will change as OpenAI develops its own search index. Google AI Mode is in active development. Perplexity's source preferences could shift as the platform grows. The two-reality framework described here is accurate as of February 2026. It may need updating by April.

If you've been advising clients on AI visibility as a unified strategy, this data suggests your advice was incomplete. So was ours. The field is younger than the confidence we've been projecting.

What makes these findings worth publishing now is not that they're definitive. It's that they converge. Multiple independent research teams — Ray, Averi.ai, Digital Bloom, and Fishkin — using different methodologies on different datasets reached findings that point to the same structural conclusion: AI search is not one thing, and treating it as one thing produces bad strategy.

---

Methodology

This analysis was produced through the M.A.R.C. methodology — Machine-assisted research, human-curated content. All sources are named, linked where possible, and evaluated against a published source quality hierarchy. Where sources conflict (the 53% vs 76.1% discrepancy, for example), we explain the difference rather than picking whichever number suits the argument. Where data is limited, we say so. Confidence levels are stated explicitly.

Findcraft provides AI visibility services. This is disclosed because you deserve to know the incentive behind the content. The analysis stands or falls on the evidence presented, not on who published it. We have no commercial relationship with any of the researchers cited in this analysis.

---

Further reading

These are independent sources — none of them are affiliated with Findcraft:

• Lily Ray — "Are Citations in AI Search Affected by Changes in Organic Search?" — The 11-site correlation study showing organic losses cascade to AI citation losses across platforms.
• Lily Ray — "Is Google Finally Cracking Down On..." — Analysis of self-promotional listicle penalties and their connection to AI visibility losses.
• Semrush — AI Overview Research — Data on AI Overview prevalence (15% of SERPs) and citation source preferences.
• Semrush — Google AI Mode — Research showing only 53% domain overlap between AI Mode citations and organic top 10. CTR impact data via Pew Research (~49% reduction).
• Onely — "Semantic SEO for AI Search" — Compilation and analysis of cross-platform citation data from Averi.ai and Digital Bloom, including structural signals and content format performance.
• Rand Fishkin / Near Media — Episode 244 — AI visibility measurement methodology (600 volunteers, 65-90 prompts per platform), narrow vs broad market behaviour.
• Search Engine Land — "Authority in AI Search" — Three-pillar authority framework: category, canonical, and distributed authority.
• Wil Reynolds / Seer Interactive — February 2026 Interview — "Mortgaging trust to get back time" — on AI content quality, authentic voice, and trust as a measurable signal.
• Barry Schwartz / Search Engine Roundtable — Weekly Recap, 20 February 2026 — AI Mode 98-language rollout, publisher opt-out exploration, review removal instability.

---

Frequently asked questions

Findcraft provides AI visibility services for London's craftspeople. This article reflects our analysis of publicly available research — including the parts where the data is limited or the picture is incomplete.

Content produced through the M.A.R.C. methodology — our framework for evidence-based, ethically-governed content.