Core Web Vitals and AI Citation Correlation: Does Page Speed Affect Citations?

TL;DR

Good Core Web Vitals will not directly buy you AI citations, but bad ones will cost you. A 2026 analysis of 107,352 AI-cited pages found CWV are not a direct ranking lever for AI Overviews; performance only matters when it is bad enough to hurt trust signals and engagement (Search Engine Land, 2026). Separate analysis of ChatGPT-cited pages found pages with FCP under 0.4 seconds averaged about 3x more citations than pages with FCP over 1.13 seconds (SE Ranking, 2026) — a strong correlation, but not a controlled causal study.

Core Web Vitals reference

INP replaced FID on March 12, 2024, and remains a confirmed Search ranking signal (Ahrefs, 2025; Google Search Central, 2025). Field data comes from CrUX (Chrome User Experience Report).

What public studies actually show

Search Engine Land 107K-page analysis (2026)

• Sample: 107,352 webpages appearing prominently in Google AI Overviews and AI Mode.
• Finding: distribution of CWV across cited pages is roughly the same as across the broader web. CWV did not predict AI Overview inclusion at the median.
• Caveat: pages in the poor CWV bucket were under-represented, suggesting a floor effect rather than a positive lift.
• Author's framing: "Core Web Vitals don't boost AI rankings — except when performance fails badly enough to hurt trust and engagement."

SE Ranking ChatGPT citation analysis (2026)

• Pages with FCP under 0.4 seconds averaged 6.7 ChatGPT citations.
• Pages with FCP over 1.13 seconds averaged 2.1.
• Same study found llms.txt adoption had negligible predictive impact on citation likelihood.

Practitioner observations

Multiple agency analyses report a CLS effect: pages with CLS in the poor range (> 0.25) are over-represented among low-citation domains (andlisten, 2025; Blue Compass, 2026). Hypothesis: layout instability suggests low-investment templating, which AI retrievers downweight via engagement and trust proxies.

Causality vs correlation

The published studies are observational. None randomizes CWV improvements and re-measures citation outcomes. Three plausible mechanisms:

1. Direct retrieval impact. Slow origins time out for AI fetchers, especially OAI-SearchBot and PerplexityBot, which do not retry aggressively. A page that consistently times out cannot be retrieved and therefore cannot be cited.
2. Engagement and dwell-time proxies. AI systems that incorporate Bing or Google index quality signals inherit downstream effects of CWV on user engagement.
3. Confounding template quality. Sites that invest in performance also invest in editorial structure, schema, and citations — the real lift may be from content quality, with CWV co-varying.

Treat reported correlations as a useful signal of which sites tend to be cited, not a guarantee that fixing CWV alone will lift citation rates.

Why AI fetches differ from search ranking signals

• No interaction simulation. ChatGPT and Claude crawlers do not run JS, so INP is not measured by them; they care about TTFB and HTML payload size.
• No CrUX equivalent. AI platforms do not have a published field-data dataset; correlations are inferred via citation studies.
• Crawler tolerance varies. Googlebot and GoogleOther retry on slow responses; PerplexityBot and OAI-SearchBot are more likely to drop the URL.

Prioritization recommendations

1. Fix anything in the poor bucket first — LCP > 4s, INP > 500ms, CLS > 0.25.
2. Drive TTFB below 600ms for AI-bot user agents (most significant for non-rendering crawlers).
3. Eliminate CLS on key article templates (reserve image and embed dimensions; defer chat widgets).
4. Then optimize LCP and INP into the good range for the human-side ranking benefit.
5. Verify with both lab tools (Lighthouse) and field data (CrUX, RUM).

Methodology disclosure for citing CWV-AI claims

When referencing third-party studies in your own content:

• State sample size, source, and date.
• Distinguish correlation from causation explicitly.
• Avoid the implied claim that fixing CWV alone will increase citations.
• Cite primary sources where possible; secondary write-ups (especially vendor blogs) may overstate effect sizes.

Common myths

• "Sub-2.5s LCP guarantees AI Overview inclusion." False. The 107K-page study showed no median lift.
• "INP doesn't matter for AI bots." Half-true — it doesn't matter for non-rendering crawlers, but it still affects human engagement signals that feed into Bing and Google indices.
• "llms.txt plus fast pages beats everything." llms.txt is currently low-signal for ChatGPT citation probability (SE Ranking, 2026).

FAQ

Q: Will improving Core Web Vitals increase my AI citation rate?

Probably not directly. Improvements into the poor bucket (e.g., reducing CLS from 0.4 to 0.1) plausibly help by removing trust and retrieval drags; improvements within the good bucket are unlikely to move citation rates measurably (Search Engine Land, 2026).

Q: Which CWV metric matters most for AI search?

For non-rendering crawlers (GPTBot, ClaudeBot, PerplexityBot), TTFB and HTML payload size matter more than INP. CLS is the most consistently correlated CWV metric in observational studies (andlisten, 2025).

Q: Does INP affect AI citation rates?

Not directly for crawlers that do not execute JavaScript. It still matters for the human-side engagement signals that feed Google and Bing indices, which in turn feed Gemini and Bing Copilot.

Q: Are these studies peer-reviewed?

No. The available evidence is industry analyses (Search Engine Land, SE Ranking, agency studies). Treat effect sizes as directional, not absolute.

Q: How should I prioritize CWV work alongside content investments?

Fix poor-bucket CWV first because it can be a citation floor. Then invest in content depth, primary sources, and structured data — these tend to drive citation rates more than further CWV optimization.