Caveats for the working-surface brief: independent anchors (Pew, peer-reviewed GEO paper, Ahrefs large-N) carry the load; vendor figures are range / corroboration, not independent confirmation

The honest summary of source quality across this brief:

• Independent anchors carry the load: Pew Research Center — 18% of Google searches in March 2025 produced an AI summary (68,879 searches across 900 US adults; independent anchor) (Pew, no commercial incentive); Aggarwal et al., "GEO: Generative Engine Optimization" (Princeton / Georgia Tech / Allen AI / IIT Delhi, KDD '24) — citations + quotations + statistics in visible text lift source visibility by >40% across queries (peer-reviewed, KDD '24); Ahrefs (16.975M citations across 7 AI platforms) — average age of AI-cited URLs 1,064 days vs 1,432 days for organic top-10: 25.7% "fresher" (large-N primary measurement, Ahrefs has vendor incentive but the data is primary).
• Vendor figures (Semrush, BrightEdge, Conductor, Amsive, AirOps, Moz) are range / direction corroboration, not independent confirmation. SEO platforms have an incentive to show that AI Overviews are big and growing.
• The Seer citation-CTR finding (Seer Interactive (53 brands, 5.47M queries, 2.43B impressions) — brand-cited pages earn ~120% more clicks per impression than uncited; ~38% behind no-AIO pages; explicit causal caveat) carries an explicit causal caveat — correlation, not proven causation.
• The interactivity / account-state capabilities (Capability 2 — interactive functionality: the visitor supplies input and the site returns a computed or looked-up result (calculator, quote, search, booking, configurator), Capability 4 — account / state: the site remembers who the visitor is and what they have done (login, saved items, order history, progress) so returning visitors resume rather than restart) lack a clean primary outcome dataset (GAP — no clean, current, primary dataset ties interactive features (calculators, booking, account state) specifically to conversion / retention for general SMB marketing websites). Argued on mechanism, not on a search-visibility lift number.
• Two widely-repeated stats are quarantined (QUARANTINE — "Interactive content drives 52.6% more engagement / 13 min vs 8.5 min" traces to a single Mediafly press release 2022; B2B sales-enablement decks, NOT marketing-website visitors, QUARANTINE — "53% of mobile users abandon a page that takes >3 seconds to load" traces to Google "Need for Mobile Speed" 2016–2017; the data was about MOBILE AD LANDING PAGES, not general business websites) — provenance issues, context mismatch, staleness.

Practical posture for Candid: lead with Pew + GEO + Ahrefs. Cite vendor trackers as a range. Disclose the Seer causal caveat. Treat schema as eligibility, not as a citation lever (GAP — no independent or primary evidence that schema markup ITSELF improves AI-answer citation; vendor claims (FAQ schema → 2.8× citations, author schema → 3×) are single-source and unverified). Argue the interactivity / account-state capabilities on mechanism grounds.

Cross-brief: Sister to the caveats umbrellas of the falling-cost-floor and searchable-catalogue briefs — same skeptical, source-incentive-flagged methodology.