A Rival Gets Cited by AI. Can You Catch Up?

Ask AI which companies in your industry are reliable, and the answer names four or five competitors. Not you. The boss looks up and asks: is there still a chance?

The conclusion first: yes, you can catch up — not by overtaking them across the board, but by being the clearest on the sub-questions they answer badly. The list AI produces is not a single ranking; it is an assembly of answers to a set of sub-questions.

The boundary too: not every position is catchable. If a competitor already holds substantial third-party endorsement and you have none at all, catching up takes time, and nobody can promise it will happen. What follows is how to raise the probability, not how to guarantee an outcome.

First, understand what "being recommended" actually is

Many people picture AI recommendations as a leaderboard — first place, second place, fixed once set. That is not how it works.

Google's documentation explains that AI Mode performs query fan-out: one user question is split into up to 16 sub-queries, each retrieved separately, then combined into a single answer.

So "which companies in this industry are reliable" may be split, at the retrieval level, into:

• which major suppliers exist in this industry;
• what to look at when judging whether a supplier is reliable;
• what the common problems are with a given product or process;
• what lead times, certifications and minimum order quantities normally look like;
• what the common pitfalls are.

Here is the crux: you do not need to win every sub-question. You need to be the clearest on a few key ones. The final answer is assembled from those sub-results — a few key pieces coming from you is enough to earn a place in the list.

Where the gap usually sits: three data points

The gap is usually in what others say about you. Research from the University of Toronto measured that when multiple sources are available, AI's preference for third-party sources reaches 53% to 95% (arXiv:2509.08919, a preprint, not peer-reviewed). Scope note: this is a range of relative preference under experimental conditions, not a fixed ratio. It suggests a competitor's lead may have little to do with a better website, and a lot to do with having more people saying good things about them than you do.

The chaser's opening is adding evidence. The KDD 2024 GEO paper (DOI: 10.1145/3637528.3671900, peer-reviewed) tested one variable: after adding source citations, quotations and statistics to content, a page previously ranked fifth gained up to 115.1% in visibility, while a page ranked first fell 30.3%. Scope note: these are relative changes under laboratory conditions on a fixed query set, not a promise of results. It does suggest one thing — adding evidence helps those further back and can even hurt those in front. That is exactly where a chaser should aim.

What decides the outcome is relevance and position, not keyword density. The SIGIR 2026 accepted paper "What Gets Cited", drawing on 252,000 controlled trials, finds the decisive factors are topical relevance and position in context, while formatting changes alone have an unstable effect. The same body of work finds keyword stuffing ineffective — even harmful — in AI answers (KDD 2024).

Three moves: benchmark, fill the gaps, build sources

Step one: benchmark — take their answers and yours apart

Do not just check whether they appear. Check which questions they appear on. The method: list the 10 to 20 questions your buyers are most likely to ask, put them to AI one by one, and record three things — who appears in the answer, which sources are cited, and how deep your name sits.

When you are done you have a map: some questions the competitor owns outright, some they answer thinly, some nobody answers well at all. That map is your plan of attack.

Step two: fill the gaps — pick three sub-questions they answer badly and be the clearest

Do not write "our company is better than our competitors" — that carries no information for AI. Write about the specific questions competitors answer vaguely and you happen to know cold.

An example of the direction: where a competitor writes "reliable quality", you can describe the yield differences across processes for this product category and the conditions under which problems appear. Where they write "fast lead times", you can describe how production scheduling actually changes with order volume. The more specific it is, the harder it is to copy — and the easier to extract.

Step three: build sources, so that others saying good things actually exists

This step is the slowest, and it is where the gap is widest. The method is not publishing at volume, but sustained, objective output on a small number of platforms chosen to match the source structure of the AI you are targeting — answers to industry questions, technical judgement, real experience.

The test is simple: delete the brand name — is the piece still worth reading? If nothing but advertising remains, it is a low-value source to AI, and a hundred more will not help.

Why not collide head-on

The instinctive move is to write whatever they wrote, only longer and with more keywords. That road leads nowhere, for three reasons:

1. Keyword stuffing backfires. KDD 2024 found keyword stuffing ineffective — even harmful — in AI answers.
2. Formatting changes alone are unstable. The SIGIR 2026 experiment shows formatting adjustments have an unstable effect on whether content gets cited — what actually counts is relevance and position.
3. You cannot outrun their accumulated lead. If a competitor already holds a large stock of third-party coverage, customer reviews and industry citations, colliding head-on over the same question is unlikely to turn in your favour quickly. Move to a question they answer badly, though, and the starting line levels again.

Three common questions

How long does catching up take?

No timeline is promised. What can be said: the lower the starting base, the easier change is to observe. If you are currently absent from every core question, closing any one of them is visible. If you are already mentioned often and merely ranked late, moving further up is much harder.

Should we write comparison articles — "A versus B, which is better"?

You can, on two conditions: compare facts only, without disparagement, and make every claim checkable. Content that runs down competitors may work briefly, but it also lowers your own credibility as a source — and underneath GEO sits one question: will AI trust you?

A competitor appears in AI answers, but customers say their quality is mediocre. Is that an advantage for us?

An opportunity, but one to handle carefully. The right move is to explain how to judge quality — which parameters matter, what to check at acceptance, where problems usually appear. Let the buyer draw the conclusion instead of drawing it for them. The first is professional content; the second is disparagement.

Where this leaves you

To pull this together: the list AI produces is not a solid block — it is assembled from answers to individual sub-questions.

So the right posture is not total superiority but finding one specific question they answer badly and you know cold, then answering it better than anyone. One such question beats ten articles saying "we are professional".

Start with one small thing: take the five most central questions in your industry, ask each to AI, and record who appears in every answer. Run it for two weeks and you can see which positions are locked and which are empty — the empty ones are your way in.

Yihang Network (Shenzhen) previously worked in cross-border B2B and Alibaba.com-related services, and has been extending into GEO (Generative Engine Optimization) since 2024. Our view: catching up is not a matter of collision, but of picking the right sub-questions.

Data and effect boundaries: the University of Toronto study cited here (arXiv:2509.08919) is a preprint and not peer-reviewed; the KDD 2024 paper (DOI: 10.1145/3637528.3671900) is a peer-reviewed conference paper, and its percentages are relative changes under laboratory conditions on a fixed query set; the conclusions of the SIGIR 2026 accepted paper "What Gets Cited" come from 252,000 controlled trials and reflect relative preference for candidate content under controlled conditions; query fan-out in Google AI Mode comes from Google's official documentation. None of the above represents the real market or the results of any specific company. The methods described raise the probability of being retrieved and cited; they do not guarantee that a company will appear in any AI answer, and they promise no timeline, competitive outcome or business result.

Sources

1. University of Toronto: arXiv:2509.08919 (preprint, not peer-reviewed)
2. Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024 (DOI: 10.1145/3637528.3671900)
3. SIGIR 2026 (accepted): "What Gets Cited", 252,000 controlled trials
4. Google Search Central: official documentation on AI Mode and query fan-out