How Google Click Signals Drive SEO Rankings and AI Answers (Cyrus Shepard)

Source: raw/newsletter-zyppy-signal-ee2ba02ac8.md — Cyrus Shepard, Signal by Zyppy. Published 2026-03-25.

Shepard synthesizes three primary sources — the US v Google antitrust trial, the Google API leak, and Google patents — into a conceptual framework for how Google collects and uses user click signals for ranking, and how those signals now also shape AI answers. The irony he opens with: AI answers reduce clicks to the open web, yet AI output relies on click data to be accurate.

Evidence class: a secondary synthesis of leaked/legal/patent primary sources. The existence of these signals is well-grounded (patents, trial testimony, leaked attributes are real); the exact usage is inferred. Shepard is explicit: “treat this post as a conceptual framework… we don’t have every nitty-gritty detail.” Distinct from the wiki’s AI-citation research cluster — this is about classical ranking mechanics plus how click data feeds AI grounding. Not a controlled study.

Key Takeaways

• Google’s “three fundamental signals” (ABC): Anchors (links), Body (text), Clicks. Click/engagement data is co-equal with links and content, per trial framing.
• Five click signals exposed in the API leak / patents:
  • impressions — a URL appearing in results; Google appears to test URLs in different positions to see which perform.
  • clicks — frequency of clicks relative to impressions.
  • badClicks — a short click (immediate bounce back to the SERP) is “indicative of a poor page.”
  • goodClicks — a long click (user stays before returning) is “indicative of a good page.”
  • lastLongestClicks — the click that was last and longest in a query session; the strongest satisfaction signal.
• User interaction data was argued to be the majority of “Google’s scale advantage” at the antitrust trial — sharing it would let competitors “mimic Google’s ranking.”
• Click signals now feed AI answers, three ways: (1) Google grounds AI answers with FastSearch / RankEmbedBERT using “70 days of search logs plus human-rater scores” — those logs contain click data and help generate the citation links; (2) AI Overviews / AI Mode summarize top-ranking results (of the query and fan-out queries), which are surfaced partly via click data; (3) other platforms like ChatGPT rely on Google results (SerpApi/Tom’s Guide reporting: Bing alone “isn’t enough to replicate Google’s accuracy at scale, especially for obscure or long-tail queries”).
• Manipulation is hard by design. NavBoost memorizes ~13 months of interaction data (stability, slow change); FastSearch uses 70 days; Google computes “squashed” values to filter spam/noise; “slicing” weights each click by position, query, device, and location. Click-spoofing can sometimes give a short-lived, risky bump.
• You can’t measure clicks directly — use proxies. Since competitor click data is invisible, “you are essentially competing against yourself.”

The Named Algorithms (Nerd Section)

• NavBoost / Glue — long-standing systems that build and re-rank results from click data; NavBoost memorizes ~13 months of user interactions.
• FastSearch — grounds AI answers using 70 days of search logs + human-rater scores.
• RankEmbedBERT — the embedding/ranking model referenced in grounding.
• Caveat Shepard flags: a patent or API endpoint existing doesn’t confirm exactly if/how Google uses it — but these concepts recur across the leak, patents, and trial testimony, which is a strong signal of importance.

The Playbook: Relevance → Usefulness → Satisfaction

Google is effectively asking three questions of the click data. Optimize each in order.

1. Prove Relevance — earn the click. Not a naive CTR play (Google calls clicks “noisy,” penalizes clickbait titles, and adjusts for position/appearance). Use the SERP snippet to signal relevance via Intent (what they searched), Promise (how you meet it), Brand (who delivers). Goal: earn good clicks by showing relevance — a high CTR without usefulness/satisfaction can attach bad signals to your site.
2. Prove Usefulness — goodClicks vs badClicks. Measured by how long a user stays before returning — but not raw dwell time. Google weights each click by query, user type, language, and country, then compares against all query-document combinations (patent US8661029B1 describes a long-clicks-to-total ratio used to “re-rank future search results”). “How tall is Mount Everest” needs seconds; “optimize investment portfolio for inflation” needs minutes. Demonstrate usefulness through engagement. (Google says “time on site” is not a ranking factor; yet many studies correlate it with rankings — Backlinko: +3 seconds on site ≈ one position higher.)
3. Prove Satisfaction — lastLongestClicks. The user clicks in, engages, gets fully answered, and doesn’t return to click another result. Earn it by anticipating the next need: answer related questions, show your evidence, provide next steps. Shepard’s favorite illustration (AJ Kohn): a local directory should display other sites’ ratings on its own page — otherwise it “chases users back to the search results.”

Measuring Clicks by Proxy

• CTR — Google Search Console (tricky: CTR swings by position, and AI-answer impressions muddy it; measure across URL-pattern sections). Tools: SEOTesting, SEOGets.
• Engagement / Time on Site — GA4 “Average engagement time per active user” (filter to organic). Competitor estimates via Semrush Traffic & Market or Similarweb. Time-on-site isn’t itself a ranking factor but feeds how Google computes other metrics.
• Satisfaction — no single metric. Track Returning Users (GA4), Conversions, and especially Brand Search — a branded query can both start and end with you (GSC now has a branded filter).

Open Questions

• Are clicks a direct ranking signal or ML/evaluation input? Genuine industry debate; there’s also evidence Google can predict clicks without live data. Either way, pages should “look like” results Google expects to satisfy users.
• How much of the API-leak attribute set is actively used vs vestigial? Unknowable from outside; Shepard’s confidence comes from cross-source recurrence, not confirmation.

Related

• AI Era — “Task Completion” and “Strong Brand” are the site-strategy expression of last-longest-clicks and brand search.
• Zyppy AI Citation Ranking Factors Meta-Analysis — Search Rank (#2) is downstream of these click signals; this article is the mechanism beneath that factor.
• Zyppy AI Citation Playbook — the “answer near the top / satisfy the user” tactics that earn goodClicks.
• SEO Content Ops Fundamentals — the wiki’s other Google-mechanics / ROI-measurement reference.
• Google’s Generative AI Search Optimization Guide — Google’s official line that AI features draw from the same Search systems these signals rank.

Try It

1. Filter GA4 engagement to organic traffic and find your lowest “average engagement time per active user” page templates — likely badClicks generators.
2. Rewrite one SERP snippet (title + meta) for Intent + Promise + Brand rather than for clicks alone; watch CTR and engagement together in GSC/GA4.
3. Add “next-step” content to a page users currently leave — related questions answered, evidence shown, other-source ratings displayed (the AJ Kohn move) — to chase the last-longest-click.
4. Start reporting Brand Search via GSC’s branded filter as a satisfaction/brand-strength KPI.