TL;DR
- When an AI answer names a competitor you consider weaker, check the order things were found in before you blame anyone's quality. In a University of Illinois preprint (arXiv 2602.03608, version 2, 21 October 2026, with no peer-review note in the version we read), four AI models were handed Amazon's top ten results for a product search and asked to rank them. With nothing changed, the tenth result reached the top five in 0 percent of trials.
- Text added to that tenth listing moved it. Written as a step-by-step rationale, or as a customer review that the researchers' software wrote, the added text put the listing in first place in 76.5 to 88.5 percent of trials across 15 product categories and four models.
- The review version read as natural. Twenty annotators judged it manipulated 18.4 percent of the time, against 12.0 percent for untouched listings, and a perplexity filter (a check on how predictable the text reads) took 5.5 to 7.5 points off its first-place rate.
- The limits are wide: English Amazon product searches, model versions from 2024 and 2025, the authors' own ranking instruction instead of a live engine, and detection tested only against weak detectors (a perplexity score, a phrase list, a length cap) plus 20 annotators. Nothing in it was measured in Quebec, in French or on a local service. Nobody judged whether the promoted product was a worse choice.
- None of the three techniques is something we offer. A review written by the seller in a customer's voice is a fake review. Canada's Competition Bureau calls the practice astroturfing, and the Competition Act prohibits representations that are false or misleading in a material respect.
- What an owner can do: check whether your own pages are found at all, read the text on the competitor's listing, keep dated captures, and report what you find instead of copying it.
A business owner asks an AI assistant for a recommendation in their own trade and gets the competitor down the road, the one with fewer years, fewer reviews and a thinner site. The usual explanation is not foul play. In what drives AI citations we covered a 90-day study in which every page ChatGPT cited was one its own search had returned, and in the rich get richer we covered what a model already remembers about a brand. This post adds a third explanation, from a paper that measured it directly: text added to a listing to steer the model that ranks it.
We wrote this post as a defence. It describes the mechanism, how large it was in the lab and what it looks like from the outside. It is not a set of instructions, and we left out the paper's prompts and settings on purpose.
The source is "Controlling Output Rankings in LLM-Based Search via Generative Engine Optimization" by Haibo Jin, Huimin Zeng, Yifeng Luo and Haohan Wang of the University of Illinois Urbana-Champaign. The authors call their method CORE. Its limits come before its figures. It is a preprint, and the benchmark is the authors' own. We read version 2 in October 2026, have not re-run any of it, and found no code or data release named in it. The searches are English Amazon product searches, run on models older than the ones in use at the end of 2026. Every figure below is the authors' unless we label it as ours.
What did the study test?
The study tested whether one seller, changing only the text of its own listing, can move that listing up a list that an AI model ranks. The authors built a benchmark they call ProductBench: 15 Amazon product categories, 200 product searches in each, and for every search the top ten results from Amazon, with the name, price, description, rating and review count of each.
Four models did the ranking, reached through their official APIs: GPT-4o (the November 2024 version), Gemini 2.5 Pro, a model the paper calls Claude-4 without naming the variant, and Grok-3. The models did not search for anything. Each one received the search query and the ten results as a list, with an instruction to rank them from most to least relevant using only the information provided. The instruction is the authors' own, not the one a live engine runs.
The target was always the tenth result. The rule of the test, in the authors' words: "The only modifiable part is the text of items shown in the search results, such as product names, descriptions, or reviews." They appended one of three kinds of text to the tenth listing:
- A character string. A short string optimized on an open model. It does not read as language.
- A step-by-step rationale. A passage that walks through the candidates the way a shopper might reason and concludes in favour of the target.
- A review. The same argument rewritten as the account of a customer who bought the product. No customer wrote it.
Software wrote the last two. That is as much of the method as this post describes. The score is the promotion success rate: the share of trials in which the tenth result landed in the top five, the top three or first place.
Does an AI model follow the order it was given?
In this experiment the last item never climbed on its own: with no text added, the tenth result reached the top five in 0.0 percent of trials, on every one of the four models and in every one of the 15 categories. The authors' explanation: "The baseline success rate is 0%, since LLM outputs largely follow the retrieval order, leaving items at the end of the list almost no chance of reaching the Top-k."
That zero is partly built into the design. The target is always the last of ten, and the score counts one thing, a climb into the top five. The paper does not report how much the models reshuffled the other nine. So the result is narrower than "the model copies the list". What it shows is that the last item did not climb on its merits.
It points the same way as the 90-day study in what drives AI citations, which used a different method, with the engines' own web search switched on. There, in the authors' replay through the ChatGPT API, the model cited 43.8 percent of the pages its own search ranked first and 3.5 percent of those ranked 11th to 20th, and all 22,315 pages it cited had come back from its own searches.
Position kept its weight even among doctored listings. In one test, on the Home & Kitchen category, the authors appended a different kind of text to each of the top three results and rotated their order. The review-style text took first rank in 71.0 to 75.0 percent of cases when its listing sat in first position. When its listing sat third, it took first rank in 26.5 to 62.0 percent of cases, depending on which of the other two sat above it.
For an owner, this is the first and most ordinary answer to the question in the title. The competitor was found ahead of you, or you were not found.
How far did added text move the tenth listing?
On the authors' benchmark, a software-written rationale or review put the tenth listing first in 76.5 to 88.5 percent of trials, and a character string did so in 30.0 to 38.5 percent. The table gives the share of trials in which the tenth result was ranked first.
| Text appended to the tenth listing | GPT-4o | Gemini 2.5 Pro | Claude-4 | Grok-3 | Lowest and highest across 15 categories |
|---|---|---|---|---|---|
| Nothing (original order) | 0.0% | 0.0% | 0.0% | 0.0% | 0.0% |
| Character string | 33.0% | 31.0% | 34.0% | 30.5% | 30.0% to 38.5% |
| Step-by-step rationale | 81.0% | 77.5% | 83.5% | 77.0% | 76.5% to 85.5% |
| Review written by the software | 79.0% | 83.5% | 82.0% | 78.5% | 77.0% to 88.5% |
The paper's headline figure (first place in 80.3 percent of trials, top five in 91.4 percent) is what the authors report as their average "under the best optimization strategy". The paper does not show how it was calculated. Four details change how to read it.
- Each model was steered by itself. The model that wrote and revised the text was the model that ranked the list. When text tuned on GPT-4o was applied unchanged to the other three models, in the Home & Kitchen category, the first-place rate fell by 14.5 to 23.5 points for the rationale and by 15.5 to 20.5 points for the review. That still left the tenth listing first in 55.0 to 68.0 percent of trials.
- The cost is small. The authors count between one and six US cents per listing for the rationale technique, so cost is not a barrier.
- It was not limited to product listings. On 200 passage-ranking queries and 200 travel-itinerary queries, the rationale put the target first in 69.0 to 81.0 percent of trials. Those tests are also in English and also in the lab.
- The scoring script is the authors' own. In GEO tricks, measured, one method's text scored about five times higher under its authors' own script than under a shared one. CORE's figures come from its authors' script only, and we have not seen them re-scored under a shared protocol.
Can a doctored listing be detected?
In the paper, the character string was caught by the perplexity check and by nearly every human reader, the rationale by most checks, and the review by few. The authors measured two things: how predictable the text reads to an older language model (perplexity, where a low score reads as ordinary language), and what 20 human annotators made of it.
| Text | Perplexity (GPT-2) | Fluency rated by annotators, 1 to 5 | Judged manipulated by annotators |
|---|---|---|---|
| Untouched listing | 30 to 31 | 4.7 | 12.0% |
| Character string | 1,488 to 1,572 | 1.2 | 98.5% |
| Step-by-step rationale | 68 to 75 | 3.7 | 62.1% |
| Review written by the software | 31 to 33 | 4.6 | 18.4% |
Read the first row before the last. Annotators suspected 12.0 percent of listings nobody had touched. The software-written review was flagged 18.4 percent of the time, 6.4 points above that false-alarm rate. Twenty annotators on five samples is a small study, and the paper does not say how many judgments stand behind each percentage. It says only that the review did not stand out to a reader.
The authors then tried three filters and measured what was left of the first-place rate.
| Filter the authors tried | Character string | Step-by-step rationale | Review written by the software |
|---|---|---|---|
| No filter | 30.5% to 35.0% | 77.0% to 82.0% | 78.5% to 83.5% |
| Perplexity threshold | 0.0% | 3.0% to 3.5% | 72.0% to 78.0% |
| List of 20 phrases typical of step-by-step reasoning | Unchanged | 14.0% to 18.5% | 56.0% to 61.0% |
| Length cap | Unchanged | 29.0% to 33.0% | 62.0% to 68.0% |
The perplexity filter removed the string entirely and nearly all of the rationale. It took 5.5 to 7.5 points off the review. The phrase list did the most against the review: 21.5 to 23.5 points. The authors' own summary: "Review still remains the strongest overall performer under defense mechanisms".
That does not make the review undetectable. The paper tested a perplexity score, a phrase list and a length cap. It did not test an AI model asked to judge whether a text is trying to manipulate a ranking. GEO-Bench, the benchmark we covered in GEO tricks, measured, ran that kind of judge on other techniques and found it separated every effective one from clean text, though a filter set at one threshold caught only some. We have not found a published result for that kind of judge on this paper's review text. What the paper establishes is narrower: the review got past a perplexity threshold and most of 20 human readers, and a 20-phrase list still cut its first-place rate by 21.5 to 23.5 points.
Is a review you wrote yourself a legitimate tactic?
No. A review written by the seller, or by the seller's software, in the voice of a customer who never bought anything is a fake review. Canada's Competition Bureau described fake reviews in 2015, in the first volume of its Deceptive Marketing Practices Digest: astroturfing is "the practice of creating commercial representations that masquerade as the authentic experiences and opinions of impartial consumers, such as fake consumer reviews and testimonials." The same text states that the misleading advertising provisions of the Competition Act "prohibit advertisers from making a representation, by any means whatever, that is false or misleading in a material respect".
We are not lawyers, and we cannot tell you whether a given text crosses that line. The business risk does not need a ruling. The technique rests on a statement the seller knows to be false, placed where customers and machines will read it as a customer's.
The paper presents CORE as a way to help the small sellers that retrieval order buries. Its own impact statement adds that "this capability exposes risks of ranking manipulation". We read the paper for that second sentence. We do not write reviews for clients, we do not add text addressed to a ranking model, and we do not offer any of the three techniques, under any name. The other two are no better: the string is spam, and the rationale is text written for the machine and placed in a field meant for the customer.
What does this not tell you?
It says nothing about Quebec, French, local services, live assistants, how often this happens, or whether the shopper was worse off.
- Anything about Quebec, French or local services. The searches are English Amazon product searches. No local query, no service business and no French text was tested. The mechanism may carry over. The percentages do not.
- How the live assistants behave today. The models were reached through their APIs, handed a list, and given the authors' instruction. Whatever a live engine does before and around ranking is outside the experiment, and the model versions date from 2024 and 2025. A later preprint with overlapping authors (arXiv 2611.01126, 31 October 2026) tested the method on live engines that were handed the pages to rank. We have not covered it here.
- Whether the promoted product was worse. No human judged the quality of the result. "Weaker" is the owner's word in our title. The paper measured movement in a ranking and nothing about harm to the shopper.
- How often this happens. The paper measures what is possible in a lab. It gives no count of listings doctored this way on any marketplace. At a few cents a listing we assume it is tried. We have not measured it.
- How a serious detector would do. Detection rests on one perplexity score, one phrase list, one length cap and 20 annotators.
- The decimals. A few appendix values disagree with the main table by a point or two. The direction is consistent, and we would not quote any cell to a client as a rate.
What should you check when a competitor is the AI's answer?
Check the ordinary explanations first, in the order below, and treat doctored text as the last one. Steps 1 to 3 come from our earlier posts. Step 4 comes from this paper.
| Check | What to look at | What it tells you |
|---|---|---|
| 1. Were you found at all? | Ask the question with your town in it, in a new conversation, several times. Read the sources each answer lists | If your pages and listings never appear among the sources, the problem sits before any ranking. See the AI visibility funnel |
| 2. Where was the competitor found? | Which of its pages or listings the answers cite, and which third-party pages name it | Being found first, or in more places, is the ordinary reason a business is named |
| 3. Does the model already know the name? | Ask the same question with web search turned off, where the assistant allows it | A name the model remembers can be recommended with no page behind it. In our July 2026 check on a Quebec trade, models answering from memory named a local business we could verify in 5 of 70 responses, so for a small local competitor this explanation is unlikely, on that one probe (one trade, medium confidence). See the layer no tactic fixes |
| 4. What does the text on their listing say? | The description fields of the listing or page the answer cites. Look for a customer's story told in the seller's own description, a walk-through that compares the alternatives and concludes for the seller, or a run of meaningless characters | One possible sign of the mechanism in this paper. It is not proof: sellers also quote real reviews in their descriptions |
| 5. Did you keep a record? | The question as typed, the date, a capture of the answer and a capture of the listing text | Without dated captures there is nothing to report and nothing to compare next month |
If step 4 turns something up, report it and leave it there. Where the marketplace or review platform has a reporting channel, use it. The Competition Bureau takes complaints about deceptive marketing through its online form. Answering in kind is the wrong move, for the reason in the section above and for two practical ones the paper supplies. The gains were tuned to one model and fell on the others. And the filters are in the platform's hands: the phrase list that cut the review's rate was written by the paper's own authors.
What we do instead
This paper changes nothing in how we work, and it explains why the first step is where it is. Our GEO work starts by measuring: we run the questions buyers ask through the engines that matter, with the town in the question and each question asked at least twice (the procedure is in our post on which city a report was measured from), and we record whether a client's pages are retrieved, cited or passed over. Page changes come after that, and they are aimed at answering those questions, as laid out in rewriting a page for AI answers. None of it guarantees a mention.
We sell this work, so we have an interest here. Test what we say the way you would test anyone's: ask what was measured, on which engine, and what was changed on your pages and why. Our Local SEO Audit, one of our SEO audits, includes a review of how a business shows up in AI answers.