A Tel Aviv company that sells AI search services published an essay on Medium on November 17, 2026, arguing that the self-ranked "best of" pages B2B software brands produced in bulk through 2024 and 2025 now supply Google's AI Overviews with the competitor names that end up recommended to buyers.
The essay, titled "I Wrote Best-Of Listicles for Years, and Google's AI Now Uses Them to Recommend My Competitors," runs under the Hordus ai account and is written in the first person by an author who is not named. Its evidence comes from a study by SEO consultant Lily Ray, published on her Substack on June 17, 2026, which tracked Google's AI answers for 100 B2B "best" queries. Its closing section compares three products that measure AI answers. One of the three is Hordus AI's own service.
In Short
Hordus AI, a company that sells AI search services, published an essay saying Google's AI Overviews take facts from brands' self-ranked "best of" pages and then recommend the rivals listed on them. Its figures come from Lily Ray's June 2026 study, in which the brand behind the page was left out of the recommendation in 224 of 323 citations, while Ray's September follow-up put that share at 83% and a separate Scrunch study found a small lift for self-ranked brands. For B2B marketers, being cited and being recommended are now separate results, and the essay that makes the point ends by comparing three products, one of them its author's own.
A formula, described from the inside
"For a long stretch of my career, I wrote the same article over and over," the essay opens. The template it describes will be familiar to anyone who has searched for software: a "10 Best Tools" piece in which "the client went first, with a glowing paragraph and a screenshot," while "the competitors went below, with two lines each and a polite mention of their pricing page."
According to the author, the pages did their job. "It worked, and I did not think much about it," the essay says. Reading Ray's research changed that view, because the format "now does something I never intended: it hands buyers to the brands listed underneath."
PPC Land has followed the ranked listicle as a format under pressure for most of 2026, and the essay places its author among a crowd. Ray's research, it notes, tracked 184 such pages from 146 brands, many of which had published more than one. "Everyone in B2B content knew the template," the author writes.
The essay is careful not to condemn the earlier logic outright. "I want to be fair to my younger self," the author writes. "In 2024 and 2025, a self-ranked listicle was a rational move." Niche B2B categories had very few independent pages comparing vendors, so a brand that published its own roundup often filled the gap, ranked in search and was picked up by early AI answers "because there was not much else to pick," according to Hordus AI.
What the research measured
The essay's numbers come from Ray's June 17 write-up, which does not appear as a link anywhere in the essay. According to Ray, she selected one canonical "best [category] software" question for each of 100 B2B categories and pulled Google's AI Overview answers at three checkpoints: April 15, May 15 and June 8, 2026. The data came from Ahrefs' Brand Radar tool. About one in five of those queries produced no AI Overview at all, leaving 80.
Ray drew a line between two outcomes. A brand's page counted as cited when it appeared among the sources of the answer, either in the text or in the sidebar. A brand counted as recommended only when the AI Overview named it as one of its picks. That distinction is the whole story.
| Measure | Figure | Note |
|---|---|---|
| B2B "best" queries tracked | 100 | One canonical query per category |
| Checkpoints | April 15, May 15 and June 8, 2026 | AI Overviews only |
| Queries that produced an AI Overview | 80 | About one in five produced none |
| Self-promoting pages identified | 184, from 146 brands | Many brands published more than one |
| Citations of brands' own lists | 323 | Across the 80 queries with an AI Overview |
| Citations where the listing brand was not recommended | 224 (69%) | Ray's headline figure |
| Citations where the listing brand was recommended | 99 (31%) | PPC Land calculation from 323 minus 224; the essay gives 31% |
| Queries citing a self-promoter's list without recommending it | 74 of 100 | 74 of the 80 queries with an AI Overview, or 92.5% (PPC Land calculation) |
Two details in the essay's retelling are compressed. It writes that "across 80 of those queries, Google cited brands' own listicles 323 times," which is consistent with Ray's data but leaves out that 80 is simply the number of queries that triggered an AI Overview. The 31% figure does not appear in Ray's write-up; it is the remainder once 69% is subtracted. Ray does not state whether the 323 citations are counted across all three checkpoints combined.
The examples are the sharpest part. For a query on the best LMS for selling courses, the essay says, "Google leaned on Oasis LMS's own listicle throughout its answer, then recommended Kajabi, Thinkific, LearnWorlds and Teachable instead." TMetric's task management roundup was cited, while Todoist, Asana, Trello and ClickUp received the recommendation. Ray's own wording on Oasis LMS is more guarded: she writes that its mentions of those competitors may have helped them. "I have written that exact page. More than once," the essay's author adds.
Cited, then passed over
Why would a system read a page and then ignore its conclusion? The essay offers an explanation, framed as an inference rather than a finding. A self-ranked list, it argues, does two jobs at once: it explains a category, and it claims the author is the best option in it. "AI Overviews appear to take the first job and ignore the second," according to Hordus AI.
On that reading, the recommendation is drawn from sources without an obvious stake in the result. In Ray's data, Reddit, Forbes and YouTube were among the domains most often cited for best-of queries. The author's conclusion is blunt. A carefully structured page about a competitor's strengths was "in effect, a briefing document for the answer that recommended that competitor."
Google has not described how AI Overviews weigh self-interest in a source. Ray herself acknowledged in June that the real mechanism is difficult to establish from outside the company.
Newer numbers, pointing in different directions
The essay cites only the June study. It does not mention that Ray published a follow-up about six weeks before it appeared. In that update, dated October 4, Ray reported that across 100 "best B2B software" queries measured in September 2026, brands that ranked themselves first were left out of the recommendation in 83% of Google AI answers citing their pages, up from 69% in June. AI Overviews were also citing 38% fewer vendor-written best-of pages than in June. In Ray's sample, then, the pattern the essay describes became more pronounced while the number of such citations fell.
Other datasets do not line up neatly behind that conclusion. Scrunch, which sells AI visibility software, analysed about 10,000 URLs cited by seven AI answer surfaces between May 17 and June 30, 2026, and found that a cited self-ranking page lifted its author's recommendation rate from about 4% to roughly 7%, while its mention rate rose from about 19% to about 39%. A competitor named on the list was recommended instead in 24.3% of those answers, and 63.9% recommended no brand at all. For AI Overviews specifically, the listing brand was recommended in 8.9% of answers that cited its page.
Kevin Indig's Growth Memo analysis, published August 24, 2026, measured the organic side and found that 62% of vendor-authored listicles had gained estimated organic traffic since January, with a median gain of 38%. The same study found that the share of listicles at position one had fallen from between 18.9% and 20.2% in January to between 15.1% and 15.6% in August, on a fixed set of 2,400 queries. AI Overviews appeared on 93.4% of B2B results pages in that dataset.
That makes three datasets built on three methods, each produced by a consultant or a vendor with clients of its own. None of them comes from Google.
The organic side: January and May
The essay extends the argument beyond AI answers. "Ray found that many domains leaning heavily on the tactic saw organic visibility fall sharply around January 2026, and some drops sped up through Google's May 2026 core update," it says. It adds a caveat in Ray's favour: "She is careful to say listicles were one of several possible causes, but it is not a risk I would take again."
PPC Land's record fills in the background. In a May 2026 interview, Ray said that dozens of companies were hit on January 20, 2026, nearly all with many listicles, and that one had published 2,000 articles each claiming first place. SISTRIX reported in February that sites built on self-promotional lists lost up to 49% of their visibility, with every identified case in SaaS. ClickUp's blog fell from 1,193,114 monthly organic visits in January 2025 to 28,790 in April 2026, a property in which every listicle placed ClickUp first.
Google then stated on May 15, 2026 that its spam policies apply to AI Overviews and AI Mode, and six days later Google began rolling out its May 2026 core update. Such updates reshuffle rankings broadly rather than penalising a single site, which is one reason attributing any one site's losses to its listicles remains guesswork.
How the rewrite process works, according to Hordus AI
The essay does not argue for abandoning comparison content. "I have not stopped writing comparison content. I have changed how I write it," the author says. According to Hordus AI, the revised approach has four steps.
- Audit before publishing. The author lists every best-of page on which the brand ranks itself first and checks whether it still brings traffic, links or conversions. The next check is whether AI Overviews cite the page while recommending someone else. Pages "doing more for competitors than for the brand get rewritten or retired."
- Criteria instead of rankings. Each option is described by who it suits, where it is strong and when the commissioning brand is the wrong pick. Writing a sentence that names a competitor as the better choice for some needs "felt strange the first time," the author writes, adding that it "made the page far more believable, to readers and to machines."
- Proof on head-to-head pages. The essay lists feature tables, pricing logic, integration lists, migration steps and customer evidence. "A claim of superiority with nothing behind it reads as marketing to everyone."
- Work outside the brand's own site. Since recommendations come from independent sources, the essay says, effort moves to review platforms, analyst coverage, industry publications, community threads and creator reviews.
The essay's FAQ repeats the first step in shorter form: pages that still earn traffic, links or conversions are kept, while those that mostly supply information about competitors are rewritten or retired, according to the document.
Three products, one of them the author's
"The lesson I keep coming back to is that being cited and being recommended are now two different outcomes. Most teams only track the first," the author writes. The essay then describes three products it calls the options the author hears about most. Its descriptions are summarised below.
| Product | Type, per the essay | AI engines named | Pricing, per the essay |
|---|---|---|---|
| Profound | Analytics platform for mid-market and enterprise teams; shows anonymized prompts and responses and which source pages are cited; needs an in-house owner | Not listed | Quote-based beyond the trial |
| Peec AI | Self-serve tool tracking visibility, position and sentiment, benchmarking competitors and identifying sources; "less granular than Profound" | ChatGPT, Perplexity, Gemini "and more, depending on the plan" | Published monthly plans |
| Hordus AI | Managed service running a detect, assess, engineer and validate cycle, then building and placing content "meant to close the gaps it finds"; GEO Site Audit scores pages from 0 to 100 | ChatGPT, Gemini, Claude, Perplexity and Google AI Overviews | Not stated |
According to the essay, Profound and Peec AI show the gap between citation and recommendation, while Hordus AI "also does the work of closing it." The comparison is written by one of the three companies in it, and Hordus AI is the only entry described as changing AI answers rather than measuring them. The last three of the essay's six FAQ entries concern Hordus AI's own service, including whether it can help rewrite self-ranked listicles. "Its process is built for that kind of gap," the answer reads.
Hordus AI describes itself on Medium as a GEO/AEO "influence company designed to map, measure, and move AI answers." The company is headquartered in Tel Aviv, according to an April 29, 2026 release announcing its GEO Site Audit, which named Dovev Goldstein as chief executive. That release said each page receives a GEO Score from 0 to 100 across four dimensions: AI Accessibility, Structured Data, Content Completeness and Internal Link Authority. The company's website also lists Grok among the engines it covers, which the essay does not mention. Its profile on Startup Nation Central's Finder database, last updated in July 2026, lists Hordus AI as founded in 2026, pre-funding, with one to ten employees.
The market for this kind of measurement has grown crowded. A day before the essay appeared, PPC Land reported that a Peec AI researcher had found ChatGPT linking to external websites in nearly half of conversations, up from about 10%, with 63% of linked URLs absent from the same chat's cited sources. Even the vendors' own data shows that "cited" covers several different things.
Why this matters for the marketing community
For B2B marketers, the practical consequence of Ray's distinction is about measurement. A dashboard that counts citations can report a page as a success while the answer built on it sends the buyer elsewhere. The essay's point, stripped of its sales section, is that the two numbers need to be read separately.
The concern is not new. In May 2025, Ray documented AI-generated lists that named companies as the best in their category being cited by AI Overviews as authoritative sources. In early 2026, the Wall Street Journal reported on businesses paying to shape ChatGPT's recommendations, with one agency saying 44% of its clients' referrals now came from AI platforms. Google's May 2026 statement on spam policies covering AI answers brought the same question inside its formal rulebook.
What has changed is the direction of the risk. The self-ranked page was once a cheap way to fill an empty category. According to Ray's June data, as retold by Hordus AI, it now risks educating the machine about competitors. According to Scrunch, it can still lift a brand modestly. According to Indig, most such pages gained organic traffic this year. Content teams deciding what to do with years of accumulated comparison pages are working from evidence that is recent, small in sample and partly produced by companies that sell the remedy.
The essay's fourth step also carries a cost it does not quantify. Moving effort to review platforms, analyst coverage and creator reviews shifts spending from owned pages, which a brand controls, to third parties, which it does not.
What the essay leaves open
The essay does not name its author, link to Ray's research or mention her September update. It gives no pricing for Hordus AI, no client results and no criteria for selecting the three products it compares. Ray's study rests on one query per category across 100 categories, and neither Ray nor the essay establishes why Google's systems separate citation from recommendation.
Key facts
- Document: "I Wrote Best-Of Listicles for Years, and Google's AI Now Uses Them to Recommend My Competitors"
- Publisher: Hordus ai account on Medium; author not named; six-minute read
- Published: November 17, 2026
- URL: https://medium.com/@hordus.ai/i-wrote-best-of-listicles-for-years-and-googles-ai-now-uses-them-to-recommend-my-competitors-e53fb61820f1
- Underlying study: Lily Ray, "Why Calling Yourself the Best Could Be Helping Your Competitors Win in AI Search," Substack, June 17, 2026 (https://lilyraynyc.substack.com/p/why-calling-yourself-the-best-could)
- Study checkpoints: April 15, May 15 and June 8, 2026; data from Ahrefs Brand Radar
- Headline figures: 100 queries; 80 with an AI Overview; 323 citations of brands' own lists; 224 (69%) without a recommendation for the listing brand; 74 of 100 queries citing a self-promoter's list without recommending it; 184 pages from 146 brands
- Follow-up: Ray, October 4, 2026: 83% left out in September; 38% fewer vendor-written best-of pages cited than in June
- Products compared: Profound, Peec AI, Hordus AI
- Company: Hordus AI (hordus.ai), Tel Aviv; chief executive Dovev Goldstein; GEO Site Audit announced April 29, 2026, with a GEO Score from 0 to 100
- Engines Hordus AI covers, per the essay: ChatGPT, Gemini, Claude, Perplexity and Google AI Overviews (its website also lists Grok)
Timeline
- 2024-2025: Self-ranked "best of" pages spread across B2B categories with few independent comparisons, according to Hordus AI.
- May 14, 2025: Ray documents AI-generated "best" lists cited as sources in Google AI Overviews.
- January 20, 2026: Dozens of list-heavy sites lose visibility, one of them holding 2,000 self-ranking articles.
- January 30, 2026: The Wall Street Journal reports on businesses paying to shape ChatGPT recommendations.
- February 2026: SISTRIX reports visibility losses of up to 49% for sites built on self-promotional lists.
- April 2026: ClickUp's blog traffic is down 97.6% from its January 2025 level.
- April 15, 2026: First checkpoint of Ray's 100-query AI Overviews study.
- April 29, 2026: Hordus AI announces its GEO Site Audit from Tel Aviv.
- May 15, 2026: Second checkpoint; Google states its spam policies apply to AI Overviews and AI Mode.
- May 21, 2026: Google begins the May 2026 core update.
- June 8, 2026: Third and final checkpoint of Ray's study.
- June 17, 2026: Ray publishes the study: 224 of 323 citations (69%) without a recommendation for the listing brand.
- August 24, 2026: Kevin Indig finds 62% of vendor listicles gained organic traffic since January.
- September 24, 2026: Scrunch finds a cited self-ranking page lifts its author's recommendation rate from about 4% to 7%.
- October 4, 2026: Ray reports 83% of AI answers citing self-ranked pages left their authors out in September.
- November 16, 2026: Peec AI data shows ChatGPT linking to websites in nearly half of chats.
- November 17, 2026: Hordus AI publishes its essay on Medium, citing Ray's June figures.
Related PPC Land coverage
- Google AI Overviews cut self-ranking listicle citations 38%, Lily Ray finds - Ray's September measurement, with the 83% figure and the drop in citations of vendor-written pages.
- Self-ranked brands gain AI recommendations from 4% to 7%, Scrunch finds - A vendor study across seven AI surfaces that points to a small lift for self-ranked brands.
- Vendor listicles gain 38% median traffic since Google's January demotion - Kevin Indig's analysis of 5.32 million organic results showing most vendor listicles gained traffic.
- Lily Ray: what the SEO industry is getting dangerously wrong about AI search - Ray's account of the January 20, 2026 losses and the company with 2,000 self-ranking articles.
- ClickUp's blog lost 97.6% of its traffic - here is exactly how - A case study of a self-ranking listicle programme and its traffic collapse.
- AI Overviews cost Germany 265M organic clicks monthly, SISTRIX data shows - SISTRIX's review, including visibility losses of up to 49% for self-promotional list sites.
- Google spam policies now officially cover AI Overviews and AI Mode in Search - Google's May 15, 2026 documentation change bringing AI answers under its spam rules.
- Google's May 2026 core update is live - and the clock is ticking - The start of the update the essay links to accelerating losses.
- Google's AI Overviews face significant spam problem - Ray's May 2025 finding that AI-generated "best" lists were cited as sources of truth.
- How brands manipulate ChatGPT to dominate AI search results - The Wall Street Journal's investigation into paid efforts to shape AI recommendations.
- ChatGPT links to websites in nearly half of chats, up from 10%, Peec AI finds - Peec AI data on the gap between links and cited sources in ChatGPT answers.
Summary
Who: Hordus AI, a Tel Aviv company that sells managed AI search services, publishing under its Medium account; SEO consultant Lily Ray, whose research supplies the essay's figures; and the B2B brands, including Oasis LMS and TMetric, whose self-ranked pages were cited in Google's AI Overviews while competitors were recommended.
What: A first-person essay arguing that self-ranked "best of" pages now feed Google's AI Overviews with competitor information, citing Ray's finding that the listing brand was left out of the recommendation in 224 of 323 citations (69%). It describes a four-step rewrite process and closes by comparing Profound, Peec AI and Hordus AI's own service.
When: The essay was published on November 17, 2026. Ray's data was collected on April 15, May 15 and June 8, 2026 and published on June 17, 2026; her September update, published October 4, put the figure at 83%.
Where: Google AI Overviews for 100 B2B "best" software queries, with the essay published on Medium.
Why: According to Hordus AI, AI Overviews appear to use self-ranked pages as a source of category facts while discounting the author's claim to be best, drawing recommendations from sources without an obvious stake such as Reddit, Forbes and YouTube. Other datasets on the same question, from Scrunch and Kevin Indig, point in different directions.
Discussion