You can rewrite every page on your website and still not get recommended. That is the part most AEO advice skips.
When a model decides which businesses to name, it is not only reading your site. It is drawing on listings, review platforms, forum threads, editorial coverage and directories, and weighing whether those independent sources agree with what you claim about yourself. If nothing outside your own domain confirms your positioning, you are asking to be taken at your word against competitors who have receipts.
Yext studied 6.8 million AI citations and found 86% came from brand-managed sources: 44% first-party websites and 42% business listings. Listings are close to half the opportunity and are usually the cheapest gap to close. Fix listing consistency first, then reviews, then earned coverage, then community presence. Never buy or fabricate any of it.
Where do AI citations actually come from?
Mostly from sources brands can influence, which is better news than it sounds.
Yext's study of 6.8 million AI citations found that 86% came from brand-managed sources, split 44% first-party websites and 42% business listings (Yext, Oct 2025).
Read that split carefully, because it reframes the work. Roughly half of what gets cited is your own site, which most AEO advice already covers. The other half is listings, which most AEO advice barely mentions. If you have optimised your website and ignored your profiles, you have done one half of the job and are competing against people doing both.
The payoff for getting into the answer is measurable. Seer Interactive found brands cited in an AI Overview earned 35% more organic clicks and 91% more paid clicks than those that were not (Seer Interactive, Sep 2025).
Why third-party sources carry weight
A model generating a recommendation is managing risk. Naming a business is an implicit endorsement, and independent corroboration is what makes that endorsement defensible.
Your own site states what you do. A listing confirms you exist, where, and in what category. A review platform indicates whether customers agree. An editorial mention shows a publisher considered you worth writing about. Those are four independent confirmations of the same claim, and their agreement is what makes a model comfortable naming you.
This also explains a common frustration. Businesses with excellent websites sometimes lose AI recommendations to weaker competitors, and the usual reason is not content quality. It is that the competitor is corroborated everywhere and they are corroborated nowhere.
| Source type | What it confirms | Effort to influence |
|---|---|---|
| Business listings | Existence, location, category, contact facts | Low, mostly administrative |
| Review platforms | Customer sentiment and service reality | Medium, needs a real process |
| Editorial coverage | Third-party judgement of relevance | High, needs something worth covering |
| Community and forums | Unprompted peer opinion | High, and cannot be forced |
| Industry directories | Category membership and credentials | Low to medium |
Start with listings, because they are half the opportunity
Listings are unglamorous and they are 42% of the citation surface. They are also the only part of this where the work is almost entirely within your control.
The goal is not merely being listed. It is being listed identically everywhere. Contradictions between profiles create exactly the ambiguity that suppresses recommendation.
Audit what exists. Search your business name and find every profile, including ones created years ago by former staff or automatically generated by data aggregators. Old profiles with a previous address actively damage entity clarity.
Reconcile every field. Name, address, phone, category, service descriptions and hours should match your website exactly. Not approximately. A model resolving conflicting facts about your category has a reason to prefer a competitor whose facts agree.
Claim and complete. Unclaimed profiles cannot be corrected, and incomplete ones give a model less to work with.
Prioritise by relevance. Your Google Business Profile, primary industry directories and the platforms your buyers actually consult matter more than volume. Fifty consistent listings on relevant platforms beat two hundred inconsistent ones.
Duplicate profiles with conflicting information. A business relocates, creates a new profile, and the old one stays live with the previous address for years. Both get crawled. The model now has two contradictory answers about where you operate, and the safest thing it can do is recommend someone else. Finding and removing duplicates is often the single highest-value hour in an off-site programme.
Reviews, and why the pattern matters more than the score
Reviews confirm that customers exist and that their experience matches your claims. For AI recommendation, the useful signal is broader than a star rating.
Recency matters, because a strong average built entirely three years ago says little about current service. Volume matters, because a handful of reviews is thin evidence either way. Distribution matters, because reviews concentrated on one platform look different from reviews across several. And specificity matters most: reviews that mention particular services, locations and outcomes give a model far more usable material than "great service, highly recommend."
The way to earn specific reviews is to ask specific questions. A request that asks which service someone used and what problem it solved produces more useful text than a generic prompt for feedback.
Do not buy reviews, write them yourself, incentivise positive-only feedback, or gate review requests so only happy customers are asked. Platform policies prohibit it, penalties are real, and a review corpus that looks manufactured undermines exactly the corroboration you are trying to build. The same applies to any third-party mention: fabricated proof is worse than no proof.
Digital PR: earning coverage that models read
Editorial mentions carry weight because a publisher chose to include you. That choice cannot be manufactured, which is precisely why it counts.
The version of digital PR that works for AI citation is narrower than general brand PR. You are not chasing awareness. You are trying to get your brand named in the context of the services you want to be recommended for, on sources a model is likely to have processed.
Original data is the most reliable route. Publishers cover findings. If you have proprietary data from your own operations, aggregate and anonymise it into something genuinely new. This works because it gives a journalist a reason to write rather than a request to.
Expert commentary is the fastest route. Journalists need qualified sources on deadline. Responding quickly and substantively to relevant queries earns named mentions, and the compounding effect of being quoted repeatedly in one subject area is exactly the association you want.
Industry publications beat general press. A trade publication your buyers read is more useful for recommendation than a general outlet with a larger audience but no topical association.
Partnerships and case studies work when they are real. Joint content with named partners, published on their domain as well as yours, creates corroboration on a source you do not control.
What does not work: syndicated press releases distributed at volume, guest posts on low-quality sites built for links, and anything that exists purely to place a mention. These produce a footprint that looks like manufactured authority rather than earned relevance.
Reddit and community presence, without getting banned
Community platforms carry real weight because they contain unprompted peer opinion, which is the hardest kind of evidence to fake and therefore the most valuable when genuine.
They are also the fastest way to damage a brand, because communities are actively hostile to marketing and moderators remove promotional accounts quickly.
| Approach | Outcome |
|---|---|
| Answering questions in your expertise with no link | Builds standing, often the only thing that works |
| Disclosing affiliation when relevant | Required by most communities, and respected |
| Dropping links in threads you did not participate in | Removed, often with a domain ban |
| Creating accounts to praise your own brand | Detected, and reputationally severe when exposed |
| Asking customers to post on your behalf | Violates most platform rules |
| Hosting an AMA where the community wants one | Legitimate and useful when genuinely invited |
The realistic posture is participation without extraction. Contribute genuinely useful answers in your area of expertise, disclose who you work for when it is relevant, and accept that most contributions will not mention your brand at all. Mentions that arrive from other people, because you were useful, are the ones that carry weight. Mentions you place yourself carry none and risk the account.
This is slow. It is also not something a competitor can buy past you, which is exactly why it holds value.
A realistic sequence
Doing all of this at once is how off-site programmes stall. Work in this order.
| Stage | Focus | Why it comes here |
|---|---|---|
| 1. Audit | Find every existing mention and profile | You cannot fix what you have not found |
| 2. Reconcile | Make listings agree with your site | Highest value per hour, 42% of citations |
| 3. Reviews | Build a recent, specific, distributed base | Confirms claims with customer evidence |
| 4. Earned coverage | Data, commentary, industry publications | Slower, compounds over time |
| 5. Community | Genuine participation in relevant places | Slowest, hardest to fake, most durable |
| 6. Measure | Track which sources get cited | Tells you where to spend next |
Stages 1 and 2 are usually a matter of weeks and produce the clearest movement. Stages 4 and 5 are ongoing programmes rather than projects, and treating them as campaigns with an end date is why most of them fail.
Measuring whether any of it worked
The measurable outcome is not mentions. It is whether the sources carrying your mentions show up in AI answers about your category.
Track a fixed prompt set covering your core buying questions, re-run it on a schedule, and record which sources each engine cites. Over time this tells you which platforms actually influence answers in your specific category, which is far more useful than a generic list of directories.
Two cautions. Model outputs vary between runs, so a single check proves very little and you need repeat runs to distinguish a real change from normal volatility. And off-site work has a lag, because a new mention has to be crawled and processed before it can influence anything. Judging a digital PR programme after three weeks will tell you nothing.
Our complete metrics framework covers prompt set design and repeat-run cadence in detail. For the on-site half of this work, see structuring content for AI search and the technical guide to schema and crawler access.
Want to know which sources are citing your competitors?
Book a free AI Visibility Audit. We will show you which third-party sources AI models use when recommending businesses in your category, and where you are missing from them.
Book Your AI Visibility AuditFrequently Asked Questions
How many business listings do I actually need?
Fewer than most directory services will sell you, and consistent rather than numerous. Your Google Business Profile, the major directories in your industry, and the platforms your buyers genuinely consult cover most of the value. Fifty consistent listings on relevant platforms outperform two hundred inconsistent ones, because contradictions between profiles actively suppress recommendation.
Do backlinks still matter for AI citations?
Links matter, but not in the way link building traditionally treats them. What appears to carry weight is the mention and the context around it, meaning your brand named alongside the services you want to be recommended for, on a source a model is likely to have processed. A link from a low-quality site built for SEO adds little. A named mention in a respected industry publication adds a lot, whether or not it carries a link.
How long does off-site work take to affect AI visibility?
Longer than on-site changes, because a new mention has to be published, crawled and processed before it can influence anything. Listing corrections tend to show up soonest since the platforms are crawled frequently. Earned coverage and community presence compound over months. Anyone promising off-site results in weeks is describing listings work and calling it digital PR.
Can I pay for placement in AI recommendations?
Not through the organic citation path, which is what this guide covers. Paid placements are appearing separately as advertising products, and our guide to ChatGPT ads and what they mean for AEO covers that. Buying reviews, mentions or manufactured coverage in an attempt to influence organic recommendation is a different matter: it violates platform policies and produces a footprint that undermines the corroboration you are trying to build.
What if my industry has no relevant directories or communities?
Then reviews and earned coverage carry proportionally more weight, and the general-purpose profiles you do have matter more. It is also worth checking the assumption, because most categories have niche directories, professional bodies, accreditation registers or trade publications that are easy to overlook precisely because they are not consumer-facing. Those often carry more weight in a specialised category than a general directory would.
Ashur Homa
Built and scaled a digital brand to $100M+ in sales with zero ad spend. Has helped businesses generate millions through AI go-to-market strategy. Leads growth at Omni Eclipse.
Connect on LinkedIn
