By Michael Semer
By Michael SemerWhat you’ll learn:
- Why good content can hurt your AI visibility when near-miss pages compete with the page that actually answers the query.
- How distractor debt creates deranking gravity, pulling retrieval toward outdated, overlapping, or inconsistent content.
- Why publishing more isn’t the fix, and how consolidation, redirects, and a clear canonical answer can improve your chances of being cited.
I’ve written before about what happens when a B2B team floods its own site with AI-generated filler and Google’s site-wide signals eventually come looking for the rest of the domain. That’s a real problem, but at least it’s an obvious one.
This problem is sneakier because there’s nothing really wrong with the content. No slop, no content mill craptent spew, no intern instructed to produce 47 “thought leadership” posts before lunch. Every page may have been written by someone who knew the subject, reviewed by someone competent, and published for a perfectly defensible reason.
And your company still can’t get the page it wants cited because all those perfectly respectable pages are invisibly fighting each other.
Here’s the short answer why
Content that’s semantically close to a buyer’s query but doesn’t quite answer it can be more damaging to AI visibility than content that’s obviously irrelevant. It competes with your best page for retrieval, and sometimes the near miss gets picked instead of the page that actually answers the question.
We can call the accumulation of this stuff distractor debt. Like technical debt, nobody puts it in the annual plan, everybody accumulates some, and eventually you discover you’re paying interest. Accumulate enough of it and you create deranking gravity: all those near misses exerting downward pressure on the page you actually want to win.
The near miss beats the direct hit more often than you’d think
There’s a term for this in retrieval research. A 2025 ACL paper by Chen Amiraz, Florin Cuconasu, Simone Filice and Zohar Karnin formalized what the authors call the distracting effect: what happens when a retrieved passage is semantically similar to the query but doesn’t contain the answer.
Take note of that distinction. Because we’re not talking about some completely unrelated page about your company picnic getting mistaken for a product definition. We’re talking about content that looks enough like an answer to get invited into the conversation with AI search, then proceeds to say the wrong thing.
Maybe the worst finding is that distracting passages can damage you even when the correct answer is sitting right there in the retrieved material. The near-miss content pulls the model away from the direct hit, and this effect showed up across different models, so this isn’t some charming little quirk of one AI engine that everyone can ignore until the next release.
Earlier work from Cuconasu and colleagues at SIGIR 2024 also looked at what happens as distracting material accumulates. Performance deteriorated as more distracting documents entered the context, with a substantial effect appearing even with very little noise.
In other words, you don’t need forty bad candidates for bot retrieval. One can do damage.
Now go look at your resource hub, but don’t look for garbage (though if you find any, by all means get rid of it). That’s too easy. Look for the good stuff that’s almost right: the webinar recap that covers 70 percent of the topic, the customer story that’s technically about implementation but spends its first three paragraphs explaining the category, the 2023 pillar page that got replaced but never redirected because it still gets traffic, or the thought-leadership article whose definition of the problem predates your current positioning by eighteen months.
Every one of those may be a legitimate asset. Every one may also look, to a retrieval system, like a plausible answer to the query you’re trying to win.
That’s distractor debt, and distractor debt creates deranking gravity.
Your pages disagree with each other and nobody has noticed
Here’s where this gets especially ugly for B2B enterprises. Retrieval systems work better when the documents they retrieve agree with one another, while actual B2B websites frequently look like the aftermath of a family argument.
Marketing calls the product a platform. Sales calls it a suite. The CEO called it a system of record on a podcast in April, while the careers page describes something else entirely because a recruiter wrote that copy in 2022 and nobody looked at it since. Meanwhile, two product pages define the customer’s core problem differently, because the positioning argument never got really internally settled and each writer wandered off with a different version of the truth.
Research into conflicting sources in search-augmented language models shows why this matters. When retrieved sources contradict each other, answer quality suffers because the model has to deal with the conflict before it can deal with the question.
Think about what that means when somebody asks an AI engine, What does this company do?
The engine isn’t simply finding your answer and repeating it. You’ve given it four answers, so now it has to arbitrate among them.
Once you’ve forced the engine to decide which version of your company is credible, you’ve also created an opening for outside sources with stronger corroboration. That might be G2, a competitor’s comparison page, or a press release from 2021 describing a company you haven’t been for three years.
You didn’t get outranked. You hung the jury.
Deranking gravity
Every near miss and obsolete page you leave sitting around is another rock in the pile: the webinar recap, the unredirected pillar page, the three articles that define your category three different ways, the old product page somebody forgot existed.
Eventually the pile becomes a planet, and planets exert gravity.
Then you commission a genuinely great page with clear positioning, direct answers, strong evidence, good structure, and everything else the SEO/AEO/GEO deck says you’re supposed to have. You launch it, but it doesn’t matter how elegant the rocket is if it never reaches escape velocity. The page arcs upward and drops right back into the same gravity well as everything you’ve been publishing since 2023.
That’s deranking gravity. The problem isn’t necessarily that there’s anything wrong with the page you’re trying to promote. It’s that your own corpus contains enough semantically similar alternatives to keep pulling retrieval away from it.
Aligned content works differently. If a hundred pages consistently reinforce what you do, who it’s for, and what problem you solve, the mass works in your favor. It’s the near misses that create the drag because, to a human, they’re adjacent content. To a retrieval system, they’re candidates.
The distinction really matters. Distractor debt is what you’ve accumulated. Deranking gravity is what that accumulation does to the page you want to win.
Why you can’t publish your way out
The standard marketing reflex is obvious: write the definitive page. Commission one enormous, beautifully researched, strategically optimized asset that settles the question once and for all.
Except it doesn’t, and the reason is mechanical rather than philosophical. Publishing asset forty-one doesn’t remove assets one through forty from the retrieval pool. You’ve simply added candidate forty-one to a collection of pages that already couldn’t agree on the answer.
If the problem is that the engine has too many plausible things to choose from, giving it one more plausible thing isn’t a strategy.
I’ve covered the tactical layer elsewhere, including some of the tricks people try to make content more attractive to AI systems. Different problem, similar conclusion: the work that actually matters is structural, and in this case structural work often means subtraction.
Pick the answer. Not “align around a narrative” or “socialize the messaging framework.” Actually decide what you want the market to understand. Get the necessary people into a room, including somebody with enough authority to end the argument, and settle it.
Then make the corpus say it.
Consolidate near-miss pages into the page that should win, 301 the pages you kill, and rewrite the pages worth keeping so they support the canonical answer instead of quietly auditioning an alternative. Where different pages genuinely need different answers because they address different intents, make those differences unmistakable.
The goal isn’t to make your website repetitive. It’s to reduce the deranking gravity by giving retrieval systems fewer plausible-but-wrong choices.
The slog that gets ranking results
This is a miserable marketing job because it produces no shiny new asset. Nobody unveils a content consolidation project at the QBR and gets a standing ovation, and there’ll be no heroic screenshot of the dashboard suddenly animating its way to glory. Well, unless you want to make one. If you do this job, you’ll deserve it.
It’ll take an afternoon, or a day, or multiple days spent reading your own back content catalogue and discovering that your company has defined the same term three different ways in four years, which is about as much fun as it sounds. Every marketer I’ve suggested this kind of work to has visibly wanted to discuss something else.
But here’s the problem: the bots have already read all of it.
They don’t care which page reflects your current thinking. They don’t know which messaging deck won the argument, and they certainly don’t know that the 2022 page is “basically deprecated” even though it’s still indexed, linked, and sitting there cheerfully explaining the wrong thing.
You do.
So if you want AI systems to understand what your company does, the job isn’t only to publish the right answer. It’s to stop publishing, preserving, and indexing all the almost-right ones that keep pulling the right one down.
Sources
- Amiraz, C., Cuconasu, F., Filice, S., & Karnin, Z. (2025).
The Distracting Effect: Understanding Irrelevant Passages in RAG.
Proceedings of ACL 2025.
ACL Anthology - Cuconasu, F., et al. (2024).
The Power of Noise: Redefining Retrieval for RAG Systems.
SIGIR 2024.
Semantic Scholar - Cattan, A., et al. (2025).
DRAGged into Conflicts: Detecting and Addressing Conflicting Sources in Search-Augmented LLMs.
arXiv - Puerto, H., Gubri, M., Green, T., Oh, S. J., & Yun, S. (2025).
C-SEO Bench: Does Conversational SEO Work?
NeurIPS 2025 Datasets and Benchmarks Track.
OpenReview
Frequently asked questions
Why does AI cite the wrong page from my website?
Often because another page on your site is a strong semantic match for the query even though it’s a worse answer. Retrieval happens before the final answer is generated, so a topically adjacent page can compete with the page you’d prefer the system to use. When enough of those alternatives accumulate, they create what I call deranking gravity. Consolidating overlapping content and making one page the clearest canonical answer reduces that ambiguity.
What is distractor debt?
Distractor debt is the accumulation of pages that are legitimate and topically relevant but compete with the page that most directly answers a query. The individual pages may be perfectly good content. The problem emerges collectively, when retrieval systems have multiple semantically similar candidates and no obvious reason to choose the one you prefer.
What is deranking gravity?
Deranking gravity is the downward pressure distractor debt puts on the page you want an AI or search system to retrieve. The more near-miss pages, obsolete definitions, and conflicting descriptions in your corpus, the harder it becomes for one canonical page to emerge as the obvious answer.
What is a distracting passage?
A distracting passage is content that’s semantically similar to a query but doesn’t contain the answer. Research presented at ACL 2025 found that these near-miss passages can interfere with retrieval-based answers even when relevant information is also available, because they look useful enough to compete for attention without contributing the correct answer.
Does inconsistent messaging affect AI search visibility?
It can. When a company’s own pages describe its category, product, or buyer problem differently, retrieval systems may encounter conflicting evidence instead of one consistent answer. That makes it harder for the system to confidently represent the company’s preferred description and creates more opportunity for outside sources to shape the answer.
How much content is too much?
There’s no magic number, and volume isn’t the problem by itself. The useful question is how many indexed pages are competing to answer substantially the same buyer questions. If several pages overlap heavily but give slightly different answers, you’ve created retrieval competition regardless of how good each page is individually