9 AI Search Myths vs 15 Million Data Points — What Fell and What Remains (2026)

Key takeaway: An Ahrefs research roundup spanning ~15 million data points (published on SEJ, Aug 2026) tested nine widely held beliefs about AI search against measurement. The results point in three directions — the shortcuts failed (llms.txt unfetched on 97% of sites, schema effects within ±5%), search rankings lost their guarantee (only 37.9% of AI-cited pages ranked top 10, down from ~76% a year earlier), and the new battleground is brand mentions (web-mention correlation 0.656–0.709 vs 0.191–0.244 for backlinks). We regroup the nine myths into three clusters and map what practice remains where the myths fell (as of 2026-08-20).

---

Three-line summary

• Three shortcut myths collapsed. llms.txt was never fetched on 97% of publishing sites, adding schema markup moved citations by less than ±5% after 30 days, and self-promotional "best of" lists that ranked the author's brand first did not secure brand mentions.
• The search–AI relationship has been rewired. 88.46% of ChatGPT citations still come from search results, but only 37.9% of cited pages rank in the top 10 — search remains the entrance, and page one no longer guarantees the citation.
• Mentions beat backlinks. Branded web mentions (0.656–0.709) and YouTube mentions (0.735–0.740) correlate with AI mentions far more strongly than backlinks (0.191–0.244). And all of this plays out in an environment where the brands appearing in answers shift 46% of the time — so a single check proves nothing, and repeated measurement is the premise.

The study — how to read 15 million data points

The roundup aggregates multiple studies Ahrefs ran between 2024 and 2026, published on Search Engine Journal on August 18, 2026 (original). Combined samples include 1.4 million ChatGPT prompts, 4 million citations, 863,000 SERPs, 137,000 sites, and 75,000 brands — roughly 15 million data points.

One caveat before reading: the samples are large, but this is single-vendor (Ahrefs) data published in sponsored form. Individual figures are best read as directional evidence rather than settled fact. Items corroborated by independent measurements — like the CTR decline below — are comparatively robust.

Cluster A — no shortcuts: llms.txt, schema, self-promotional lists

The most consequential result is that all three "just do this one thing" shortcuts failed under measurement.

Myth	Sample	Measured result
llms.txt drives AI visibility	137K sites, server logs	Never fetched on 97% of publishing sites; of the 3% fetched, 77% of accesses came from non-AI tools (SEO auditors etc.)
Schema markup shortcuts citations	1,885 pages with schema vs 4,000 controls	After 30 days: ChatGPT +2.2% · AI Overviews −4.6% · AI Mode +2.4% — no meaningful improvement
Self-ranked "best of" lists lift your brand	750 prompts · 34 lists · 9,886 answers	"Best" lists made up 43.8% of cited page types, but self-ranking first did not secure mentions — competitors were frequently recommended instead

The llms.txt result deserves attention. Publication spread to 28% of sites, yet the logs show AI bots simply do not come to read it. That matches the priority we laid out in our Google AI Mode optimization article: llms.txt is a supporting layer, and crawlable content exposure comes first. Schema is similar — structured data retains other value (rich results, entity disambiguation), so there is no reason to remove it, but expecting it to shortcut AI citations contradicts the data.

Cluster B — rankings are the entrance, not the guarantee

The second cluster concerns search rankings and AI citations. Two seemingly opposite findings hold at once.

88.46% of ChatGPT citations originated from traditional search results. Yet only 37.9% of AI-cited pages ranked in the top 10, 31.2% ranked 11–100, and 31.0% ranked outside the top 100. — Ahrefs research, via SEJ

The reading: AI answer engines still use the search index as their warehouse (88.46%) — a page that search cannot find rarely becomes a citation candidate. But what gets picked from the warehouse diverges sharply from rankings. The top-10 share of citations fell from roughly 76% a year earlier to 37.9%; nearly two-thirds of citations now come from beyond page one, and a full third from beyond position 100.

The practical implication cuts both ways. Search fundamentals — indexing, crawlability — remain a precondition, but "we're on page one, so we'll show up in AI" is a dead assumption. Conversely, pages with modest rankings have a real shot at AI citations — if they are structured for extraction at the question level. We covered that selection mechanism in One Query Becomes 12 Searches.

Cluster C — the new battleground: mentions over links, trends over snapshots

The third cluster sets the forward priority. In the 75,000-brand correlation analysis, the signal that moved most closely with AI mentions was not backlinks.

Signal	Correlation with AI mentions
YouTube mentions	0.735–0.740
Branded web mentions	0.656–0.709
Domain Rating (DR)	0.266–0.326
Backlinks	0.191–0.244

Correlation is not causation, so "dump backlinks, buy mentions" is premature. But the direction is clear: how often AI engines mention a brand moves with how often the brand is named across the web, far more than with its link graph. Unlinked mentions become AI-visibility assets, and the weight is shifting from link-centric SEO thinking toward accumulating brand-category co-mentions.

The volatility measurement defines how any of this work gets verified. Tracking 43,000 keywords 16+ times over a month, answer wording changed 70%, the brands shown shifted 46%, and sources swapped 45.5% — while semantic similarity held at 0.95 out of 1, meaning the substance stayed stable. The surface churns while the substance holds, so a single-day screenshot is coincidence, not measurement. When brand presence changes nearly half the time, only the trend across repeated measurements of the same questions makes a reliable diagnostic base.

Two supporting checks — traffic scale and CTR

Two remaining myths matter for practical judgment. First, ChatGPT's search-like usage was estimated at roughly 12% of Google's search volume, and across analyzed sites Google drove 39.98% of traffic versus 0.21% for ChatGPT — AI search is the growth axis, but the belief that it has already replaced search is exaggerated at current absolute scale. Second, AI Overviews reduced position-1 clicks by 58% (up from 34.5% in early 2025) — an item corroborated by Seer Interactive (65.2%+) and Authoritas (47.5%). Together the picture is clear: traffic still lives in search, but the click value of that search is being absorbed by AI answers. The unit of exposure is moving from the click to the in-answer mention.

What remains where the myths fell — diagnosis and measurement

One conclusion runs through all nine: every shortcut that reduced AI visibility to a single file (llms.txt), a single markup (schema), or a single ranking (page one) failed under measurement. What remains is brands diagnosing their own position and measuring it repeatedly — which questions mention you, which pages earn the citations, and how the gap against competitors moves.

The measurement practice comes down to three steps: (1) confirm your content is exposed in crawlable HTML before anything else (see AI Crawlers Don't Render JavaScript); (2) measure mentions and citations repeatedly on a fixed question set to extract the trend beneath the volatility; (3) track the mention axis — where your brand gets named — alongside links. Measurement tools such as RanketAI offer AI Brand Visibility Analysis, which repeatedly measures mentions and citations on your key questions, and Competitor Comparison, which measures your share gap against competing brands across the same question space.

Frequently asked questions

We already created an llms.txt — should we delete it?

No need. It costs almost nothing, and some tools do access it. But with 97% of publishing sites never seeing a fetch, placing llms.txt at the center of your AI-visibility plan contradicts the data. Crawlable content exposure comes first.

Is schema markup pointless now?

As a shortcut to AI citations, no effect was confirmed (within ±5% after 30 days). But schema retains other value — rich results, entity clarity — so there is no reason to remove it. Only the priority "schema first, for AI citations" needs correcting.

Do backlinks no longer matter?

Their search-ranking value is not what this research challenges. What it shows is that their correlation with AI mentions (0.191–0.244) is far weaker than mentions (0.656–0.709). Keep links as a search fundamental, and treat unlinked mentions — being named in reviews, communities, and media — as a separate asset worth accumulating for AI visibility.

If answers change 46% of the time, isn't measurement pointless?

The opposite. High surface volatility is exactly why a one-off check is meaningless and only repeated measurement is informative. As the 0.95 semantic similarity shows, the substance is stable — so a measured trend separates real position changes from volatility noise.

Our rankings are modest — can we still earn AI citations?

The possibility is confirmed: 31.0% of AI-cited pages ranked outside the top 100. The conditions are being indexed at all (88.46% of citations come via search) and having self-contained, extractable passages at the question level. Competing on question coverage rather than rank now has measured support.

Related reading

• Google AI Mode Optimization — 7 Signals That Matter More Than llms.txt — the crawl and indexing fundamentals that precede llms.txt
• One Query Becomes 12 Searches — Sub-Query Coverage for AI Search Visibility — the structure that selects pages by sub-query coverage, not rank
• AI Crawlers Don't Render JavaScript — the precondition before measuring: crawlable content exposure
• AI Citation Decay — Why Citations Fade Once Earned — keeping citations as an asset in a volatile environment