# Answer engines wrote rules for their ads. The citations underneath have none

URL: https://www.thedeepfeed.ai/posts/2026-10-16-manufactured-citations/
Category: Products
Published: 2026-10-16
Author: the-deep-feed
Tags: answer-engines, perplexity, openai, google, advertising, citations, seo, privacy
Kind: deep

> Perplexity cites a network of 210,000 machine-made software rankings that disclose commissions to human visitors, while OpenAI and Google build ad, tracking and payment systems on the same answers, and the only written rules cover the labeled ad slot.

## TL;DR

- **Trellner Research** counted **215,128** machine-made 'best software' pages on three linked sites and found Perplexity citing them in **41 of 380** categories. Surfer's separate sample of about **10.37 million** consumer Perplexity citations puts the same network at **64,004**, with worldmetrics.org ranked **#10**, ahead of G2.
- No one has shown these pages change what the engines recommend. The audit says so itself, and it comes from the kind of site it describes: **trellner.com** was registered on **September 1**, one day before publication, alongside sibling report brands submitted to Hacker News by one account.
- **OpenAI** promises that ads 'do not influence the answers', and its ad business reportedly hit a **$1 billion** annualized run rate. A researcher traced a one-year **__obi** cookie, filed by OpenAI under analytics, to **936** advertiser pixels on **1,029** hostnames.
- The written rules cover the labeled ad slot. The pages underneath the answers disclose commissions and paid placements to human visitors, and that disclosure disappears when an engine turns them into a footnote.

On September 2, a report titled ["Three sites made 215,128 'best software' pages for AI. Perplexity cites them"](https://trellner.com/reports/manufactured-sources-behind-ai-recommendations/) reached 519 points on [Hacker News](https://news.ycombinator.com/item?id=49536375). The claim was specific. A firm called Trellner Research asked Perplexity's two API models for a ranked top five in 380 software categories, kept every URL the models retrieved, and found three sites (wifitalents.com, worldmetrics.org and gitnux.org) that share a registrar, a Cloudflare nameserver pair and a page template, publish 215,128 pages at addresses of the form `/best/<category>-software/`, and were cited 181 times across 41 categories. Two of them title their homepages "Facts & Grounding Page." The report observed that those titles are addressed to the software that reads them.

Registry records show trellner.com was registered through NameCheap on September 1, 2026, the day before the report went up. So was hausresearch.com, whose own Perplexity audit reached Hacker News ten minutes before Trellner's, from the same account. One commenter put the question directly:

> I don't even know if I disagree with this post, but this seems really astroturfed. Why are there two anti-Perplexity articles from independent research firms with identical websites on the front-page of HN right now, submitted by the same person? Am I going crazy?
>
> — [nightpool on Hacker News](https://news.ycombinator.com/item?id=49540466), Sep 2, 2026

The starting hypothesis for this piece was that answer engines inherited the incentive problem of search optimization in compressed form: a handful of manufactured pages can steer recommendations for whole categories, while the engines build advertising and tracking businesses on the same surface. The record from the past eight weeks supports part of that and contradicts part of it. The manufactured pages are real, large and heavily cited, and an independent dataset from the consumer product confirms it. Nobody has shown they steer the answers, the audit's authors included. The engines, meanwhile, have written careful rules for the one part of the answer that carries a "Sponsored" label, and are building measurement plumbing around it that resembles Meta's. The commercial pressure their rules do not reach sits one layer down, in the citations: pages that disclose commissions and paid placements to the person who visits them, and lose that disclosure when an engine turns them into a footnote.

*Disclosure: The Deep Feed monitors whether AI answer engines cite its own pages. It is not a party to any of the studies or programs discussed here.*

# What the Trellner audit counted, and what it says it did not

Set the provenance question aside for a moment. Trellner queried `perplexity/sonar` and `perplexity/sonar-pro` through OpenRouter, one prompt per category per model, 760 calls in all. Categories were fixed before any results came back. The calls produced 3,800 recommendation slots naming 1,807 distinct products, backed by 7,534 citations to 2,055 domains. Every cited domain was checked against the Tranco popularity list and the Wayback Machine.

The headline numbers concern the long tail. In the [report's tables](https://trellner.com/reports/manufactured-sources-behind-ai-recommendations/), 59.8% of citations went to domains outside the Tranco top 100,000, and 23.4% to domains outside the top million altogether. The median rank of the ranked domains was 71,611. Wikipedia was cited three times in 7,534. The most-cited domains were familiar review and community sites (g2.com 291 times, reddit.com 261, gartner.com 158), with one exception in third place: guideflow.com, a software vendor's content blog, cited 194 times across 96 categories.

The three-site network sits further down, and the numbers are smaller than the headline suggests:

| Domain | Citations | Share of 7,534 | Categories | Registered (RDAP) |
|---|---:|---:|---:|---|
| wifitalents.com | 71 | 0.94% | 27 | Apr 17, 2024 |
| worldmetrics.org | 60 | 0.80% | 22 | May 1, 2024 |
| gitnux.org | 50 | 0.66% | 23 | Dec 4, 2023 |
| **Combined** | **181** | **2.4%** | **41 of 380** | |

The registration dates in the last column come from this publication's own registry lookups, which match the report's description: all three registered through NameCheap between December 2023 and May 2024, all delegated to the same pair of Cloudflare nameservers. The authors call this "strong circumstantial evidence" of a common account rather than proof of ownership, and note that none of the three names an owner. They also found template residue that reads like an unfilled variable: bylines reading "Within the next 26 days" and "Within the next 40 days." Worldmetrics advertises custom research "from €5,000" and vendor selection "from €2,500."

The limitations section is unusually frank. The two Perplexity tiers share one retrieval stack, and their citation lists were byte-identical in 289 of 380 categories. The categories are the authors' own, not a sample of real buyer queries, and it is one prompt, one run, one day, through the API. And the sentence that matters most for anyone citing the report: "We have not shown that any of this changes the answers." The report did not test whether removing the manufactured pages would change a single recommendation, and it did not measure ChatGPT, Gemini, Copilot or Google's AI Mode at all.

# A second dataset, from the product people use

A citation count from one day of API calls would be easy to dismiss. Stronger corroboration came from a company with no apparent connection to Trellner. On October 6, Surfer published [the 50 most-cited domains in Perplexity](https://surferseo.com/blog/most-cited-domains-perplexity/) for September 1 to 28: 57,327 tracked prompts, 978,402 answers and about 10.37 million citations, scraped from Perplexity as users see it rather than from the API. Surfer sells AI-search analytics, and it says its data skews toward marketing, SEO and Polish domains because it comes from its own customers.

The same network is in the top 50, and higher than in Trellner's sample:

| Rank | Domain | Citations (Sep 1-28) |
|---:|---|---:|
| 10 | worldmetrics.org | 19,990 |
| 11 | gitnux.org | 19,584 |
| 12 | g2.com | 19,437 |
| 24 | wifitalents.com | 13,168 |
| 32 | zipdo.co | 11,262 |

Zipdo.co is the fourth brand Trellner flagged as sharing the same nameserver pair, and its homepage, at the time of writing, carries the same "Facts & Grounding Page" title as Worldmetrics. Together the four domains drew 64,004 citations in Surfer's sample. That is 6.9% of the citations going to the top 50 domains and roughly 0.6% of all 10.37 million. Surfer's own gloss is that stats aggregators suit an engine that backs every claim with a link: "When the answer needs a number, a page with a list of dated numbers is easy to quote."

Two other names in Surfer's top 50 belong to the same story. Guideflow.com, the vendor blog that ranked third in Trellner's sample, is 23rd here with 13,392 citations. At 35th, with 10,975, is tryprofound.com, the site of Profound, which sells marketers a platform to see and act on what customers ask AI. A company in the business of AI visibility is itself one of the most-cited sources in the engine its customers want to appear in.

This publication ran one further check at the time of writing: a recount of the three sites' sitemaps. Deduplicated, they list 210,526 URLs matching `/best/…-software/`, split 70,314 (Worldmetrics), 69,751 (Gitnux) and 70,461 (Wifitalents). That is about 2% fewer than the 215,128 Trellner counted six weeks earlier. Each site lists about 110,000 URLs in total, including roughly 15,800 "best service" pages. The scale claim holds.

![Two bar panels comparing the network's citation share in Trellner's API sample and Surfer's consumer sample, with G2 as a reference bar.](/post-images/2026-10-16-manufactured-citations/two-samples-one-network.jpg)

# The audit came from the kind of site it audits

None of that settles where the Trellner report came from. Registry data for the domains the same Hacker News account submitted between August 21 and October 15 shows a pattern:

| Report brand | Registered | First HN submission |
|---|---|---|
| exhibitorlens.com | Aug 19, 2026 | Aug 21 |
| hausresearch.com | Sep 1, 2026 | Sep 2, 13:49 UTC |
| fentner.com | Sep 1, 2026 | Sep 2 |
| trellner.com | Sep 1, 2026 | Sep 2, 13:59 UTC |
| installmap.com | Sep 3, 2026 | Sep 23 |
| fundingwatcher.com | Sep 4, 2026 | Sep 15 |
| professionlens.com | Sep 15, 2026 | Sep 27 |

All seven were registered through NameCheap on the registrar's own nameservers. The brands also point at each other. Trellner's [about page](https://trellner.com/about/) names three editors (Konrad Trellner, Margaret Ellis and Sofia Marek) and says the firm "has a commercial interest in several products and datasets that appear in the magazine, including FundingWatcher, Installmap, JournalistLabs, Podcastquery and ProfessionLens." FundingWatcher's [about page](https://fundingwatcher.com/about/) says it "is a research publication of Haus Research, an independent research firm." Haus Research published the other Perplexity audit posted that day. Fentner's about page uses the same three-editor masthead and the same disclosure formula as Trellner's. FundingWatcher's report index lists four reports dated September 27 alone.

Hacker News commenters flagged the prose as machine-written and the firm as newly minted; one quoted the Trellner homepage line "The record a machine reads is not the one a company writes" and [wrote](https://news.ycombinator.com/item?id=49537187) that "this low-effort spam is exactly what this report warns about." The story stayed on the front page anyway.

This does not make the numbers wrong. Every part of the report that can be checked from outside has held up: the registry records, the sitemap scale and, through Surfer, the citation prominence in the consumer product. It does make the report an instance of the economics it describes. A brand with no history, a methodology section and a quotable number is the format answer engines cite, and the format that travels through human channels. Within a month the figure had been reworked into headlines such as ["Citation Farms: 215,128 Fake Pages Are Inside Your AI's Answers"](https://www.secureinseconds.com/blog/2026-10-01-ai-citation-farms-manufactured-sources), a claim the report itself declines to make. The number travelled. The caveat did not.

The sibling audit deserves the same treatment. [Haus Research](https://hausresearch.com/reports/perplexity-citation-audit/) asked Perplexity 310 factual questions about 210 hand-picked technology companies and checked whether each cited page contained any figure from the sentence it supported. On that deterministic test, 34.7% of 1,826 figure-bearing citations failed; scored per claim, 14.4% of 872 claims did. Only 1.3% of cited URLs were dead. Its conclusion, "The claim and the citation were produced by the same process, and that process is not retrieval," is consistent with earlier independent work: the DeepTRACE audit this publication [covered in April](/posts/2026-04-30-deep-research-agent-wars/) found 18% to 30% of deep-research citations problematic. It comes from the same cluster of new brands, and should carry the same asterisk.

# What a best-list page carries, and what an answer drops

What is on the pages engines keep citing? This publication opened the "client tracking software" page on all three sites at the time of writing.

Each is built for extraction. The Worldmetrics version is titled "Top 10 Best Client Tracking Software – 2026 Buyer's Guide," carries three bylines (written by, edited by, fact-checked by), says "Independently tested" and "Updated August 11, 2026," and embeds three blocks of structured data: an Article, a ten-item ranked list, and an FAQ page with prewritten question-and-answer pairs. The meta description is a one-sentence verdict naming Monday.com. The Gitnux page adds a scoring rubric ("Features 40% · Ease 30% · Value 30%") and the badges "AI-verified · Expert reviewed." The template residue Trellner found is still there on all three, now reading "Within the next 36 days," "Within the next 25 days" and "Within the next 40 days."

The rankings disagree. Worldmetrics and Wifitalents put Monday.com first; Gitnux puts Capsule first and Monday.com nowhere in its top five. Each page discloses a commercial relationship to the human reader. Worldmetrics: "Includes paid placements · ranking is editorial. Worldmetrics may earn a commission through links on this page." Gitnux: "Gitnux may earn a commission through links on this page — this does not influence rankings." Wifitalents: "Wifitalents may earn a commission from links on this page. This does not affect our rankings."

Disclosures like these are the norm on affiliate review sites, and they are why a careful reader discounts such pages. When an engine cites the page under a recommendation, the reader sees a domain name in a footnote. Nothing in either dataset measures how often an answer's ranking follows the cited page's ranking, which is the question that would turn this from a sourcing story into a steering story.

Vendors in the Trellner thread described the sales side of the arrangement. One, writing about an unnamed niche:

> Some of those ”best software sites” has reached out to us with an offer where we can then pay them an annual fee depending on which position we would like.
>
> — [8384727747478 on Hacker News](https://news.ycombinator.com/item?id=49538506), Sep 2, 2026

That is one anonymous account naming no site, though it matches the "paid placements" language on the page itself.

Research on what moves citations points the same way. The paper that named the field, [GEO: Generative Engine Optimization](https://arxiv.org/abs/2311.09735), accepted at KDD 2024, reported that optimization methods "can boost visibility by up to 40% in generative engine responses." A May 2026 study, [What Gets Cited](https://arxiv.org/abs/2605.25517), ran 252,000 two-source trials across six models with brand names anonymized and found topical relevance and list position were the biggest drivers of being cited first; explicit prices and recent timestamps helped reliably; formatting-only edits did little. A ranked list per category, an "Updated" date and structured data naming the winner fit those findings closely.

The search industry already has a rule for this pattern. Google's [spam policies](https://developers.google.com/search/docs/essentials/spam-policies), last updated August 28, define scaled content abuse as "when many pages are generated for the primary purpose of manipulating search rankings and not helping users," and list among the examples "Creating multiple sites with the intent of hiding the scaled nature of the content." That rule governs ranking in Google web search. This publication found no comparable published rule from Perplexity or OpenAI governing which pages their answers may cite. One commenter described the gap from the agent side:

> I do think models currently don't have enough source skepticism.
>
> If you look at agent traces when asked to compare two options to help inform a decision, many of the comparison pages cited in research are often hosted by one of the companies being compared; nearly all are AI-generated AEO plays. Not deeply considering the motive of published information is currently a glitch that can be exploited, but the window will close.
>
> — [toddmorey on Hacker News](https://news.ycombinator.com/item?id=49538500), Sep 2, 2026

![Schematic of a citation supply chain from vendor fee to best-list page to retrieval to answer, with the paid-placement disclosure dropped.](/post-images/2026-10-16-manufactured-citations/citation-supply-chain.jpg)

# The ad slot arrived with a rulebook

OpenAI's advertising business has grown on a published set of commitments. On January 16, in [a post signed by Fidji Simo](https://openai.com/index/our-approach-to-advertising-and-expanding-access/), the company listed five ads principles. Two are load-bearing here. "Answer independence: Ads do not influence the answers ChatGPT gives you. Answers are optimized based on what's most helpful to you. Ads are always separate and clearly labeled." And "Conversation privacy: We keep your conversations with ChatGPT private from advertisers, and we never sell your data to advertisers."

The timeline since then, from OpenAI's [testing post](https://openai.com/index/testing-ads-in-chatgpt/) and its updates:

| Date (2026) | Step |
|---|---|
| Feb 9 | Ads test begins in the U.S. for logged-in adults on the Free and Go tiers |
| Mar 26 | OpenAI reports "no impact on consumer trust metrics"; pilots planned for Canada, Australia, New Zealand |
| Aug 11 | Ads launched in the UK, Mexico, Brazil, Japan and South Korea |
| Sep 10 | Amazon's demand-side platform starts selling ChatGPT inventory in a U.S. pilot |
| Sep 16 | Sponsored Agents test; HubSpot and Shopify integrations |
| Oct 5 | Visual ads during image generation; expanded conversion measurement |

The testing post explains how an ad is chosen: "we decide which ad to show by matching ads submitted by advertisers with the topic of your conversation, your past chats, and past interactions with ads." [Marketing Dive](https://www.marketingdive.com/news/amazon-pilots-ad-services-in-chatgpt-what-marketers-need-to-know/829945/) reported in September that ChatGPT ads had reached a $1 billion annualized revenue run rate in late August, that OpenAI says tens of thousands of advertisers use the service, and that OpenAI controls all ad delivery decisions even for inventory bought through Amazon.

The September launch pushed the format closer to the answer. [Sponsored Agents](https://openai.com/index/reimagining-advertising-with-ai/) let a user who clicks an ad continue in "a clearly labeled conversation with a business-sponsored agent in ChatGPT." OpenAI says that conversation "is distinct from ChatGPT's independent answers and separate from the original conversation." The same post introduced opt-in "AI-powered text customization," which adapts an advertiser's headlines "to better fit the context of a conversation." A Hacker News commenter set the launch against OpenAI's help-center wording:

> this is what openai said before: "Ads do not influence ChatGPT's answers. Ads run on separate systems from our chat model, and advertisers have no ability to shape, rank, or alter ChatGPT's responses. Ads are separate and clearly labeled."
>
> — [sigmar on Hacker News](https://news.ycombinator.com/item?id=49727552), Sep 16, 2026

On October 5, OpenAI [said](https://openai.com/index/new-chatgpt-ads-format-and-measurement/) ChatGPT reaches 1.2 billion people each week, announced a visual ad format to be tested during image generation, and expanded measurement: integrations with Hightouch, Tealium and LiveRamp "to send conversion data from their existing systems to ChatGPT Ads," and Conversions API and click-attribution support with ten attribution vendors, among them AppsFlyer, Adjust and Branch.

Read closely, the principles are precise about two things: the model's answer and the ad unit. They are silent on the documents the answer is built from, including whether an advertiser can pay a third-party page that the model then retrieves and cites. That gap is not of OpenAI's making; Google Search has lived with it for two decades. What changes in an answer engine is the compression. A results page shows ten links, each with its own domain and its own disclosure. A recommendation shows one ranked list in the engine's voice, with the sources reduced to superscripts.

# The pixel that crosses sites

The tracking evidence is more detailed than the citation evidence, though it rests largely on one researcher. In April, the security researcher who publishes as Buchodi [documented](https://www.buchodi.com/how-chatgpt-serves-ads-heres-the-full-attribution-loop/) how ChatGPT ads arrive inside the response stream as `single_advertiser_ad_unit` objects, and how OpenAI's measurement script on a merchant's site stores a click token in a 30-day cookie. On September 20 the same researcher published a follow-up, ["ChatGPT now knows what you do on other websites via ad collector"](https://www.buchodi.com/chatgpt-now-knows-what-you-do-on-other-websites-via-ad-collector/), which reached 764 points on Hacker News.

> OpenAI ad collector bridges between what users are doing online and their ChatGPT profile
>
> — [@lmbbuchodi](https://x.com/lmbbuchodi/status/2101703283173663132), Sep 20, 2026

The mechanism, as described: ChatGPT requests a short-lived signed token that binds the user's account (or an anonymous device identifier) to a value called `obi`. The browser posts that token to `bzr.openai.com`, which the researcher says stands for "bazaar," OpenAI's internal name for its ads platform. The collector sets a cookie named `__obi` on the whole `.openai.com` domain with a one-year lifetime and `SameSite=None`, which lets the browser send it on requests from other sites. When a user later visits an advertiser running OpenAI's pixel, the pixel's requests carry the cookie back.

The researcher's observations from Chrome on Android:

- One `__obi` value reached 12 commercial sites under 13 pixel IDs, including Chewy, Wayfair and ThriftBooks, plus Eventbrite, HelloFresh, Coursera and SeatGeek.
- Across broader traffic, 936 distinct advertiser pixels on 1,029 hostnames; one `__obi` value appeared under ten advertisers.
- The script collects identity from advertiser-supplied fields, form fields, rendered page text and tag-manager events. Email, phone and names are hashed with SHA-256; country, region, city and postal code are sent in clear text. Page paths in the observed traffic included a medical condition, a debt-solutions funnel and a litigation intake form.
- Of 932 decoded tokens, 196 were anonymous, so signed-out users were covered too.

The caveats are the researcher's own. Testing was on Chrome for Android only; Safari's tracking prevention blocks the cookie, and so does Chrome on iOS. Only about one ChatGPT session in five produced a sync token. Most important, the researcher did not observe the server-side join that would attach advertiser-site activity to a specific account, and inferred the link from the design. A Hacker News commenter [noted](https://news.ycombinator.com/item?id=49779819) that browsers with strict cookie partitioning defeat the mechanism. Brendan Eich, chief executive of Brave, a browser that competes with Chrome, said the same and pressed the open question:

> Crucial q Buchodi asked, unanswered AFAIK:
>
> Chrome/Edge/Vivaldi/etc ChatGPT users who opt out of "marketing performance cookies" but not "analytics cookies" get navigation tracking across advertisers' sites correlatable to their chat logs.
>
> Is this okay?
>
> — [@BrendanEich](https://x.com/BrendanEich/status/2109034893711962277), Oct 10, 2026

The consent question is checkable. OpenAI's [cookie policy](https://openai.com/policies/cookie-policy/), in archived copies from September 20 and September 23, lists `__obi` under "Analytics cookies," defined as cookies that "help us understand how our Services perform and are used," with a one-year duration on chatgpt.com and openai.com. A separate category, "Marketing Performance cookies," covers measuring "the performance of marketing campaigns." The researcher reports that every decoded token carried the consent value `analytics_allowed`. A support reply to the researcher's September 14 questions acknowledged them without answering. [TechWyse](https://www.techwyse.com/news/ai-advertising/openai-chatgpt-ad-tracking-cookie-obi-cross-site) reported that OpenAI "did not respond on the substance of the findings." This publication found no substantive OpenAI statement since.

None of this touches answer independence. It does sit awkwardly beside the privacy principle, which is phrased as one-directional: conversations are kept private *from* advertisers. The pixel moves data the other way, from advertisers' pages toward OpenAI. The principles do not address that direction. The researcher's own summary of the novelty: "The mechanism is standard adtech. What has no precedent is running it on an AI chat product."

![Flow diagram of the __obi cookie from ChatGPT through a signed token and collector domain to advertiser pages carrying an OpenAI pixel.](/post-images/2026-10-16-manufactured-citations/obi-cookie-flow.jpg)

# Google's version: a bigger answer, and a price on contribution

Google runs the largest answer surface, and its moves over the same weeks went two ways. The answer grew. On August 28, [The Verge](https://www.theverge.com/tech/986364/google-search-ai-overviews-auto-expand) reported that Google had begun automatically expanding AI Overviews for some queries, placing the full summary and an "Ask anything" box above the list of links. Google spokesperson Jennifer Kutz said: "For some queries, AI Overviews may dynamically expand for topics where our systems determine it's most useful for people." A widely shared complaint about Google that month, ["When did Google get so weird?"](https://sancho.bearblog.dev/google-weird/) (2,017 points and 1,125 comments on Hacker News), described a search for old posts about a basketball player that produced an AI Overview consoling the user as if spurned by a man named Dario. "Google has seriously lost the plot," the author wrote. Ads have been inside that answer box for two years: Google brought them to AI Overviews on U.S. mobile in October 2024 and began testing them in AI Mode in May 2025, [per TechCrunch](https://techcrunch.com/2025/05/21/google-is-bringing-ads-to-ai-mode/), labeled "Sponsored."

Google also put a price on being used. [Search Engine Roundtable](https://seroundtable.com/google-al-contribution-pilot-42076.html) reported on September 14 that a small "AI contribution" pilot lets publishers "earn money when your content contributes significantly to the creation of AI-generated responses," with earnings from AI Overviews, AI Mode and the Gemini app shown in a Search Console widget. Content that only confirms facts, or is linked after a response is generated, does not qualify. [The Information](https://www.theinformation.com/articles/google-paying-100-digital-publishers-ai-overviews) reported on September 30 that about 100 publishers are being paid and that payouts vary widely. Google calls it "an early-stage learning pilot."

A payment keyed to contribution prices the act of shaping an answer, as opposed to appearing next to one. Whether it rewards the kind of page the Trellner and Surfer data describe depends on how Google judges "contributes significantly," and Google has not published that. Publishers in the pilot cannot see which pages earned money.

The answer layer also behaves differently from the link layer on price. [Productrise](https://productrise.app/blog/google-ai-mode-prefers-more-expensive-products), which sells product-visibility tracking, compared more than 2 million listings from August 9 to 31 and found that when the same product appeared in both AI Mode and the classic results on the same day, AI Mode's lead offer averaged 21.6% more expensive. The study does not report its matched sample size and pools dollars with pounds, so it is a direction, not a measurement.

# Where the hypothesis holds and where it breaks

What holds:

- Manufactured, template-built pages are cited at scale. One network of four domains drew 64,004 citations in a month of consumer Perplexity answers, and its leading domain outranked G2.
- The pages carry the features research identifies as citation levers: ranked lists, fresh dates, structured verdicts.
- The engines are building advertising and measurement on the same surface: OpenAI's reported $1 billion ad run rate, its Amazon and conversion-data partnerships, Google's ads inside AI Overviews and its contribution payments.
- The written rules cover the labeled ad slot and the conversation, not the citation layer or data arriving from third-party pixels.

What breaks:

- "A few pages steer recommendations for whole categories" is not supported. The network accounts for 2.4% of citations in one sample and about 0.6% in the other, and appears in roughly one category in nine in the API test. No study published in the window tests whether removing those pages changes an answer.
- Citation and influence may not be the same thing. The Haus audit, whatever its provenance, suggests a third of numeric citations do not contain the cited number, which would mean the footnote and the answer can come apart. That cuts both ways: a manufactured page can be cited without steering anything, and a page can steer through training data without ever being cited.
- The engines that matter most were not measured. Neither audit covered ChatGPT search, Gemini or Google's AI Mode. Surfer's data is Perplexity only.

The test that would settle the steering question is simple to state: for fixed categories, compare each answer's ranking with each cited page's ranking, across engines and days, and count how often the answer's first pick matches a page that discloses paid placements. Neither September report ran it, and this publication has not either.

![Layered schematic of an AI answer: a labeled sponsored slot governed by written principles above an unlabeled citation layer with no rule.](/post-images/2026-10-16-manufactured-citations/two-layers-one-answer.jpg)

# The disclosure that does not travel

A reader who lands on Worldmetrics' client-tracking page sees "Includes paid placements" near the top and can weigh the ranking accordingly. A reader who asks an answer engine for the best client-tracking software sees a confident list and a row of small numbers, one of which may lead to that page. The ad beside the answer gets a "Sponsored" label and a published principle; the paid placement inside the answer's evidence gets neither once it becomes a footnote. Search optimization had the same incentive problem, but a results page handed the reader ten pages to judge. An answer engine does the judging and shows its work in superscripts.

Readers adjust. Rafal Cymerys described in August [becoming "AI-blind"](https://cymerys.com/w/im-becoming-ai-blind), learning to skip low-effort machine text the way an earlier generation learned to skip banner ads. The answer engines are betting their consumer businesses on readers not developing the same reflex toward the answer box. The rules they have written so far protect the part of the answer readers already know to discount, and leave the part they are asked to trust to whoever publishes the most quotable page.

## Sources

- [Trellner Research — Three sites made 215,128 'best software' pages for AI. Perplexity cites them (Sep 2, 2026)](https://trellner.com/reports/manufactured-sources-behind-ai-recommendations/)
- [Surfer — The 50 most-cited domains in Perplexity, September 2026 (updated Oct 6, 2026)](https://surferseo.com/blog/most-cited-domains-perplexity/)
- [Haus Research — A third of Perplexity's citations don't contain the number they're cited for (Sep 2, 2026)](https://hausresearch.com/reports/perplexity-citation-audit/)
- [Trellner Research — About page (commercial-interest disclosure)](https://trellner.com/about/)
- [FundingWatcher — About page](https://fundingwatcher.com/about/)
- [Hacker News — discussion of the Trellner report (Sep 2, 2026)](https://news.ycombinator.com/item?id=49536375)
- [Buchodi's Threat Intel — ChatGPT now knows what you do on other websites via ad collector (Sep 20, 2026)](https://www.buchodi.com/chatgpt-now-knows-what-you-do-on-other-websites-via-ad-collector/)
- [Buchodi's Threat Intel — How ChatGPT serves ads: the full attribution loop (Apr 28, 2026)](https://www.buchodi.com/how-chatgpt-serves-ads-heres-the-full-attribution-loop/)
- [OpenAI — Our approach to advertising and expanding access to ChatGPT (Jan 16, 2026)](https://openai.com/index/our-approach-to-advertising-and-expanding-access/)
- [OpenAI — Testing ads in ChatGPT (Feb 9, 2026; updated Aug 11, 2026)](https://openai.com/index/testing-ads-in-chatgpt/)
- [OpenAI — Reimagining advertising with AI (Sep 16, 2026)](https://openai.com/index/reimagining-advertising-with-ai/)
- [OpenAI — Building advertising for the way people use AI (Oct 5, 2026)](https://openai.com/index/new-chatgpt-ads-format-and-measurement/)
- [OpenAI — Cookie policy](https://openai.com/policies/cookie-policy/)
- [Marketing Dive — Amazon pilots ad services in ChatGPT (Sep 10, 2026)](https://www.marketingdive.com/news/amazon-pilots-ad-services-in-chatgpt-what-marketers-need-to-know/829945/)
- [TechWyse — OpenAI ad cookie tracks ChatGPT users across sites (Sep 24, 2026)](https://www.techwyse.com/news/ai-advertising/openai-chatgpt-ad-tracking-cookie-obi-cross-site)
- [The Verge — Google further buries search results under AI mode (Aug 28, 2026)](https://www.theverge.com/tech/986364/google-search-ai-overviews-auto-expand)
- [Search Engine Roundtable — Google AI contribution pilot pays publishers for use of content in AI (Sep 14, 2026)](https://seroundtable.com/google-al-contribution-pilot-42076.html)
- [The Information — Google paying about 100 digital publishers for AI Overviews contributions (Sep 30, 2026)](https://www.theinformation.com/articles/google-paying-100-digital-publishers-ai-overviews)
- [TechCrunch — Google is bringing ads to AI Mode (May 21, 2025)](https://techcrunch.com/2025/05/21/google-is-bringing-ads-to-ai-mode/)
- [Productrise — Google AI Mode shows the same products 21.6% more expensive than traditional search (Sep 1, 2026)](https://productrise.app/blog/google-ai-mode-prefers-more-expensive-products)
- [Google Search Central — Spam policies for Google web search (updated Aug 28, 2026)](https://developers.google.com/search/docs/essentials/spam-policies)
- [Aggarwal et al. — GEO: Generative Engine Optimization (KDD 2024)](https://arxiv.org/abs/2311.09735)
- [Vishwakarma, Kumar, Jamidar — What Gets Cited: Competitive GEO in AI Answer Engines (May 2026)](https://arxiv.org/abs/2605.25517)
- [Sancho Panza's Thoughts — When did Google get so weird? (Sep 27, 2026)](https://sancho.bearblog.dev/google-weird/)
- [Rafal Cymerys — I'm becoming AI-blind (Aug 2026)](https://cymerys.com/w/im-becoming-ai-blind)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-10-16-manufactured-citations/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt