They Lied About AI Content (This Proves It)

Author: Caleb Ulku | Published: 2026-09-08 | Source: transcript (19:34, 16.9K views at ingest)

Summary

Caleb Ulku relays a study by Lily Ray and builds a local-SEO argument on top of it. The primary is now in the wiki as It Works Until It Doesn't: AI Content Strategies That Backfire; checked against it, his numbers and her disclaimers are relayed accurately. Cite the figures from the primary.

The study, as he describes it: Ray tracked more than 220 websites pulled from the customer-story pages of more than a dozen AI content platforms, so the sample is the vendors’ own highlight reel. She tracked organic traffic and organic page counts in Ahrefs and cross-checked a second dataset (ASR “Cyrix”, almost certainly Sistrix). He is explicit about the limits: these are third-party estimates, not Search Console data; Ray is not claiming the tools caused anything; and she named zero vendors and zero domains. The pattern: organic pages climb for 6–12 months of AI publishing, traffic peaks 3–6 months after the content peaks, then falls, and most of the gain is gone within the following year. He reports 54% of sites following the pattern lost 30% or more of peak traffic, 39% lost half, and 22% lost three quarters, with many ending below their pre-AI baseline. He credits Glenn Gabe with naming the chart shape “Mount AI.” Many vendor case studies are still live while the pages they describe have been removed or redirected; some brands cut their footprint in 2025–2026 and saw traffic start to return.

The delay explanation is Ray’s opinion, which he expands. Ray writes that scaled templates become “an easily detectable footprint by Google when enough sites are publishing similar pages”; Caleb’s expansion: an AI content tool is “a template and a prompt” sold to thousands of sites, so the pages share structure and shape. One site is undetectable; ten thousand form a “signature,” and Google acts once the pattern is obvious and cheap to act on. Neither offers evidence for the mechanism beyond the timing, and Ray claims no causation.

Ray found eight templates recurring in the top-traffic URLs of declining sites: comparison pages at scale, “what is X” glossary pages (sometimes machine-translated), “best X for Y” listicles, self-promotional listicles, competitor-alternative pages, programmatic location and language pages, FAQ farms, and off-topic content at scale. He stresses that every one of them works, which is why enough sites built them to create a footprint.

Two apply to local. FAQ farms (one People-Also-Ask question per URL) share the footprint problem, and he says Google deprecated FAQ rich results “earlier this year,” leaving “a few hundred thin URLs on a site that should have 40.” Programmatic location pages (one service template spun across every city in a 50-mile radius) fail, he argues, not because of volume but because each page claims a presence that does not exist. He distinguishes his own Core 30: bounded by the Google Business Profile, about 30 URLs, every one targeting the city the GBP is addressed in.

The test he attributes to Ray: could a competitor publish a near-identical page tomorrow with the same prompt and methodology? He pairs it with Information Gain and with Google’s Scaled Content Abuse policy, which he correctly says is authorship-neutral. His local angle: service businesses generate new information daily (job costs by neighborhood, recurring failures by house age, permit timing), and almost none reaches the site. AI should do structure, briefs, titles, Internal Linking (Nav vs Editorial) plans, Schema Markup (Structured Data), and research; the business supplies what only it knows. He closes by pitching the Core30 AI Tool, which he says collects Census data, Reddit threads, Search Console queries, and local-forum questions before writing, and tracks stories so none repeats. He teases his Claude's Watermarks Just Broke SEO video.

Key Claims

Ray’s sample: 220+ sites from 12+ AI content vendors’ own customer stories; Ahrefs plus a second dataset; third-party estimates; no causation claimed; no vendors or domains named. (Caleb relaying Ray; confirmed in It Works Until It Doesn't: AI Content Strategies That Backfire)
“Mount AI” pattern: pages up for 6–12 months, traffic peaks 3–6 months later, then collapses within a year. 54% lost ≥30% of peak, 39% lost half, 22% lost three quarters. (Caleb relaying Ray)
Case studies outlive the pages: vendor write-ups remain while the referenced pages are removed or redirected; some brands that pruned in 2025–2026 saw traffic return. (Caleb relaying Ray)
Footprint theory of the delay: shared tool + shared prompt produces a detectable signature once enough sites run it. (Ray’s stated belief, expanded by Caleb; not tested by the data)
Eight at-risk templates, all of which rank and get AI-cited, which is why they became a footprint. (Caleb relaying Ray)
FAQ farms: same-questions footprint, and FAQ rich results deprecated “earlier this year.” (Caleb; corroborated by Ray, who on 2026-05-13 says Google announced it “just last week”; no official Google source in this wiki yet)
Programmatic city pages fail because they assert an untrue presence, not because of page count; Google knows the real location from the GBP. (Caleb)
Core 30 is different because it is GBP-bounded: ~30 URLs, all targeting the GBP’s own city. (Caleb, defending his product)
Reproducibility test plus Information Gain: a page anyone could regenerate from the same prompt “has no reason to exist.” (test attributed to Ray)
Google history: September 2023 helpful content update; March 2024 core update with a stated 45% reduction goal for unhelpful, unoriginal content; Scaled Content Abuse policy written alongside, authorship-neutral. (Caleb; the policy wording is corroborated by Spam policies for Google web search)
Sites that survived 2023–2024 “went narrow instead of wide”: fewer pages, tighter topics, real experience. (Caleb relaying Ray’s consulting experience)
Core 30 agent collects before it writes: Census Bureau data, Reddit threads, Search Console, local forums, client differentiators, story tracking. (vendor claim)

Notable quotes

“Could a competitor publish a near identical version of this page tomorrow using the same prompt, the same methodology? That’s the test.”

Why citable: A one-line pre-publish gate that does not depend on whether AI wrote the page. Attributed by Caleb to Ray.

“Not one of those is on there because it doesn’t work. Every single one of those works. That’s the entire reason 10,000 sites are running those templates.”

Why citable: The counter-intuitive core. The risk is ubiquity, not ineffectiveness, which is why early results mislead.

“A page that says plumber in Sugar Land is telling Google and telling a customer that there’s a plumbing business in Sugar Land. No address there, no coverage there, no jobs completed there, and the page is a claim with nothing behind it.”

Why citable: Reframes programmatic city pages as a truthfulness problem, which lines up with the official doorway-abuse wording in Spam policies for Google web search. (ASR cleaned: “plumber and Sugarland”.)

“These tools work when an experienced person is holding them and they fail when someone sets them and forgets it.”

Why citable: Usable in client conversations about AI content tools. It is Caleb’s paraphrase, not Ray’s wording; she writes that the tools “should be used and overseen by experienced SEO professionals” and warns against “set it and forget it.”

Connections

Entities mentioned: Caleb Ulku, Lily Ray, Glenn Gabe, Google, Ahrefs, Google Business Profile, Core30 AI Tool, Reddit, Anthropic, Claude · Sistrix (ASR “Cyrix”; no page), G2 (minor) Concepts referenced: Scaled Content Abuse (NEW), Information Gain (NEW), Programmatic SEO, Core 30, GBP–Website Alignment (Consistency Signals), Schema Markup (Structured Data), Internal Linking (Nav vs Editorial), 8-Pass AI Writing Pipeline, GEO / AEO (Getting Recommended by AI), Local SEO, Duplicate Content (for Service-Area Pages)

Contradictions / Tensions

vs. his own Programmatic SEO advice. This will rank you #1 AND bring you customers (a.k.a. 'I Spent 10+ Years Ranking Businesses #1. This Ranks the FASTEST') and This AI Tool Will Make You RANK FAST (Full Walkthrough) promote scaled neighborhood and landmark pages, and the Core 30 course says to switch to geographic content once coverage hits 20–40% green. Here he says every Core 30 URL targets only the GBP’s own city and that spinning service pages across cities is “the fastest way I know to get a local site in real trouble.” He does not mention his neighborhood or landmark pages in this video. Possible reconciliation (not stated by him): landmark pages sit inside the real service area, city-swap pages do not. Record as an unresolved tension.
vs. James Dooley and Joy Hawkins on Duplicate Content (for Service-Area Pages). Dooley argues copied service-area pages outrank rewritten ones. Caleb now says multiplied city pages are the most dangerous local tactic. The contested status stands; this adds weight to the “don’t multiply” side, from a creator with a product to sell.
Corroborated by official Google. Spam policies for Google web search defines scaled content abuse as production-method-neutral and lists doorway abuse including multiple city or region pages that funnel to one page. Caleb’s description of the policy is accurate. Google Search's guidance about AI-generated content says the same about intent.
vs. his own 8-Pass AI Writing Pipeline and single-prompt lessons. He criticizes “one prompt in and one page out” as the architecture of the failed tools. His course teaches single-prompt drafting (Single Prompt Content Creation), though it calls the result “80% of a first draft,” not publishable copy. Partial tension only.
The footprint mechanism is untested. It is Ray’s own explanation (It Works Until It Doesn't: AI Content Strategies That Backfire), which Caleb expands, but her data does not test it and she claims no causation at all. (An earlier version of this page wrongly called it Caleb’s own theory.)
The local-SEO layer is Caleb’s, not Ray’s. Ray never mentions GBP, People Also Ask, or local service businesses; her location examples are state, country, and language pages.
Sample bias cuts both ways. Sites chosen from vendor success stories are the vendors’ best cases, which makes the declines striking, but it is not a random sample of AI-assisted sites and says nothing about AI content used with heavy human input.
Commercial. The remedy he lands on is the Core 30 agent he sells. Same flag as This AI Tool Will Make You RANK FAST (Full Walkthrough) and Claude's Watermarks Just Broke SEO.
Date wobble. He says the scaled content abuse policy came “2 years before most of these AI content tools existed”; March 2024 to this video is about 2.5 years, and AI content tools clearly existed by 2024. Treat as loose rhetoric.

Notes

Backfill done 2026-09-21: Ray’s original is It Works Until It Doesn't: AI Content Strategies That Backfire. Still worth sourcing: Glenn Gabe’s “Mount AI” writing, Google’s FAQ rich result deprecation notice (early May 2026 per Ray), and Google’s March 2024 core update announcement for the 45% figure.

He says Anthropic announced Claude watermarking “in mid August”; Claude's Watermarks Just Broke SEO (published 2026-08-14) says the mark is on models launched since 2 August. Consistent enough; not independently verified in this wiki.

No entity page for Sistrix (named only as a corroborating dataset).