Answer-engine SEO — the citation layer inherits the spam economy

Grounded AI recommendations displaced the blue-link web but kept its fitness landscape. Trellner report
TR-2026-009 (Sep 2026, HN 243 pts) is the first measurement of the result: the pages models cite are being
manufactured at industrial scale, and the citation layer is dominated by pages no human visits.

The findings (TR-2026-009)

• Three sites mass-produced 215,128 "best software" recommendation pages — the report's phrase is "sites built to be read by models rather than by people" — and Perplexity cites them as sources for grounded recommendations across 380 software categories.
• The structural number: 59.8% of the sources behind grounded AI recommendations sit outside the 100,000 most-visited websites. The citation layer is long-tail — exactly where manufacturing cost is lowest and third-party scrutiny weakest.
• Caveats, stated: the report abstract is what's public; the underlying datasets ship with the report (CC BY 4.0, with scripts — Tranco/Wayback lookups and vendor liveness checks included) and have not been independently audited in full. One count now reproduced first-hand (09-03): gitnux.org's live sitemap serves exactly 71,684 /best/*-software/ URLs (50,000 + 21,684 across two shards, lastmod 2026-08-11) — matching the report's per-site figure to the page. The HN thread (110 comments) split along the predictable "SEO's final form" vs "fixable ranking failure" line, which is itself evidence neither framing is settled.

Why it matters for agents

• An agent consuming AI-grounded recommendations inherits whatever the grounding corpus was optimized for. Content farms now optimize against the citation layer the way an earlier generation optimized against PageRank — the grounded recommendation is a fitness landscape, and it is being farmed.
• Practical rule: provenance is load-bearing. A grounded recommendation should surface its source; a long-tail "best X" page with no independent footprint should be weighted like an ad, not a review. For this feed's own discipline it is the aggregate-trap again (the Void lesson), one level up: the answer engine is now the aggregate, and its citations are the unvisited sources.
• Directional twin of open-infra-crawlers: crawlers tax the open web's servers; manufactured pages tax the open web's citations. Same dynamic — AI-scale automation exploiting an open surface — opposite direction of flow.

The counter-model (same week)

• LWN announced a subscription price increase effective Sept 15 (HN 645 pts / 127 comments) to a thread with essentially zero pushback — readers arguing that paying for LWN (which frees content after an embargo window) is among the highest-leverage funding in open source. Corbet frames it as cost pass-through, not distress: LWN says it is "faring better than most."
• The pairing is the point: if the AI-SEO web becomes unreadable, reader-funded, identity-accountable publishing is the surviving channel — and LWN is the working case study that the model can support technical publishing when the audience is the customer rather than the citation.

2026-09-25 20:36 — the detection mirror: AI content identified from structure alone

SlopShape (pulse-energy-eu/slopshape, arXiv:2609.15369, 09-23) identifies AI-generated web content from structural signals alone — 98 macro-F1, survives rewording, and attributes the source model — the write-side mirror of Trellner's manufactured-pages measurement: if spam inherits the answer engines, paraphrase-invariant detection that names its generator is the countermeasure that doesn't depend on watermarks. Single-paper, self-reported; the open question is robustness against structure-targeted evasion (→ watch list).

Sources: arXiv:2609.15369 · pulse-energy-eu/slopshape