AEO basics · · By Gerald De Oleo

How to Get AI Recommendations: Why Citations Aren't Enough

The short answer: a citation means an AI model used your page as a source. A recommendation means it told the buyer to consider your product. The first is about your content. The second is about your company — whether the model understands what you sell, who it’s for, and why it fits the buyer’s situation better than the alternatives.

Most B2B SaaS companies work on how to appear in ChatGPT answers and assume recommendations follow. They often don’t.

This post answers the questions that come up most when a company is visible in AI answers but still missing from the shortlist.

What’s the difference between a mention, a citation and a recommendation?

There are three levels, and they’re easy to confuse.

A mention means the model named your company. It might be in a list of ten, in passing, or in the wrong context.

A citation means the model used your page as a source. It pulled a fact, a definition or a statistic and credited you for it.

A recommendation means the model looked at the buyer’s question — their company size, their industry, their constraints — and named your product as an option worth evaluating.

People searching how to get mentioned by ChatGPT or how to get cited by AI are usually after the third level without realizing it. A mention or a citation is visibility. A recommendation is a verdict, and it’s the only one of the three that reliably puts you on a shortlist.

You can have many citations and zero recommendations. It happens when your blog explains a topic well, so the model quotes it, but nothing on the web makes it clear that you sell a solution to that problem. The model treats you as a reference, not a vendor. I describe this gap as the difference between being cited and being understood.

Why doesn’t a citation turn into a recommendation?

Because the model is answering a different question each time.

When a buyer asks “what is revenue attribution”, the model looks for the clearest explanation. Your educational content can win that.

When the buyer asks “what’s the best revenue attribution tool for a 50-person B2B team using HubSpot”, the model looks for something else: which products exist in that category, which of them fit those constraints, and which the rest of the web agrees are good options.

Your explainer doesn’t answer that second question. And the second question is the one at the bottom of the funnel.

The shift from being ranked to being selected is the core of AEO vs SEO. Recommendations are where that shift hits pipeline most directly.

Is “how to rank in ChatGPT” the right question?

Not quite, and the difference matters.

ChatGPT doesn’t have rankings. There’s no page two, no position eight. It writes an answer, and your company is either in it or not — and if it is, it’s described a certain way.

So when a team asks how to rank in ChatGPT, the useful translation is: when my buyer asks the questions they ask before choosing, does ChatGPT recommend us, and does it describe us correctly?

That reframe changes what you work on. Ranking is about pages. Recommendations are about whether the model has enough consistent evidence about your company to reach a verdict.

What does an AI model need before it recommends a B2B SaaS product?

In the audits I run, recommendations depend on five things. Missing any one of them is usually enough to drop out of the shortlist.

1. A clear category. The model needs to know what you are before it can suggest you. If your homepage says “the operating system for modern teams”, the model has to guess. If it says “project management software for agencies”, it doesn’t.

2. Explicit fit. Models match products to constraints: team size, industry, tech stack, budget, compliance needs. If you never say who you’re built for, the model can’t match you to the buyer asking.

3. Consistency across sources. Your site, directory profiles, review listings and third-party comparisons should describe you the same way. When they disagree, the model either picks the version with more volume behind it or leaves you out.

4. Stated differentiators. A model can’t infer why you’re better for a specific case. It needs to read it somewhere. “Built for teams without a dedicated RevOps person” is something a model can reason with. “Powerful and intuitive” is not.

5. Current evidence. Pricing, integrations, certifications, customer results. If the most detailed information about you is two years old, the model reasons from two-year-old facts.

The five requirements above apply to all of them. What differs is where each engine looks.

ChatGPT answers many product questions from what it already learned during training, and searches the web when the question needs current information. Your long-term footprint — how you’ve been described across the web over time — carries real weight.

Perplexity searches on almost every question and shows its sources. It tends to reward pages that answer the exact question directly, with specifics it can quote.

Gemini, Google’s AI Overviews and AI Mode draw from Google’s index. Here your SEO fundamentals matter most, because a page that isn’t indexed can’t be used.

Claude, when connected to search, behaves similarly: it favors sources that answer precisely and agree with each other.

The practical conclusion: don’t treat “AI” as a single channel. Track each engine separately. I explain why the engines diverge in AEO vs GEO.

What content actually gets a company recommended?

Content that answers bottom-of-funnel questions, written so a model can extract a verdict from it.

“Best for” pages. One page per segment you serve: [Product] for agencies, [Product] for fintech teams. Each one states the use case, the constraints you handle, and what customers in that segment achieved.

Honest comparison pages. [You] vs [competitor], with a clear statement of when each one is the better choice. Models trust comparisons that admit trade-offs more than ones that claim to win on everything.

Alternatives pages. Buyers constantly ask “what are alternatives to [incumbent]”. If you don’t have a page that answers it, someone else’s page will — and they choose which alternatives to list.

Public pricing and plan details. “Contact us” gives the model nothing to reason with when the buyer mentions a budget.

Integration and requirements pages. When a buyer says “it has to work with Salesforce” or “it has to be SOC 2 compliant”, the model needs a page that confirms it.

Customer results with numbers. Specific outcomes, attributed to a type of company. Models use these as evidence of fit.

Each of these answers a question a buyer asks right before choosing. That’s what separates them from content that only gets you cited.

Does it matter what other websites say about my company?

Yes, and for recommendations it often matters more than your own site.

When a model decides whether to recommend you, it looks for agreement. If your site says you’re a revenue intelligence platform but your directory listings still say “sales email tool”, the model sees two different companies. It either picks the description with more sources behind it or drops you from the answer.

The fix isn’t glamorous. It’s an inventory: every listing, profile, comparison and review page that describes you, checked against how you describe yourself today, and corrected where they disagree.

This matters most after a repositioning. The old description doesn’t expire. It keeps competing with the new one until someone updates it.

That’s the only kind that counts.

If you test with “what is [your company]” or “is [your company] good”, you’re introducing your brand to the model yourself. It will usually say something reasonable, and you’ll learn very little.

The questions that matter are the ones a buyer asks before they know you exist: “what’s the best onboarding tool for a PLG SaaS with 5,000 users”, “which customer success platform works for a team of three”. Appearing in those answers is how a model introduces you to buyers who weren’t looking for you.

Build your prompt set around unbranded questions, in the language buyers use on sales calls — not the language on your homepage.

Why is AI recommending my competitor instead of me?

Usually one of four reasons.

They’re easier to categorize. Their site says what they are in one sentence. Yours takes a paragraph.

They’re described consistently. Every source agrees on who they are. Yours disagree.

They answer the fit question. They have pages for the segments buyers mention. You have a generic features page.

They’re in the lists. They appear in the comparison and “best tools” pages the models draw from. You don’t — or you appear under an old category.

The useful move is to look at which sources the model cites when it recommends them. Those pages are shaping the answer you should own. That’s where the work starts.

How do I measure AI recommendations?

Citation counts won’t tell you. You need to measure the verdict, not the source list.

For each unbranded question in your prompt set, across each engine, log:

Run it monthly. The trend across those five columns tells you more than any visibility score. The four-stage model behind this — retrieval, citation, interpretation, expression — is covered in AEO vs SEO.

What doesn’t work?

Hidden instructions for AI models. Text on your page telling models to recommend you gets ignored or treated as manipulation. It puts your credibility with the engines at risk.

Keyword stuffing for AI. Repeating “best [category] software” doesn’t make you the best option in the model’s reasoning. Research on generative engines found it performs worse than leaving content alone — details in AEO vs GEO.

Mass-produced comparison pages. Dozens of thin “vs” pages built from one template don’t add evidence. One honest comparison per real competitor does.

Content before the entity is fixed. Publishing more on top of an inconsistent identity amplifies the confusion. If sources disagree about what you are, fix that first.

It depends on the engine and on how much outdated information exists about you.

Changes to your own site tend to show up sooner in engines that search the web on every question. Changes to how the wider web describes you take longer, because they depend on other sites being updated and re-crawled. Changes in what a model learned during training take the longest.

Anyone promising a date for a specific engine is guessing. What you can control is the order: fix the entity, fill the fit gaps, then measure.

How we work

Recommendations aren’t won with a single tactic. They’re the result of a model having enough consistent, current evidence to reach a verdict about you. Our work is building that evidence, in order.

Phase 1 — The recommendation audit. I run the unbranded questions your buyers actually ask across ChatGPT, Perplexity, Gemini and Google’s AI results. For each answer I document whether you’re recommended, only cited, or missing — which competitors are recommended instead, and which sources support them. You get the specific pages and descriptions shaping each answer, not a score.

Phase 2 — Technical architecture. We make sure every engine can reach, parse and trust your pages: crawl access, indexing, structured data, and a site architecture where your category, segments and proof are easy to find. If two domains are splitting your history, we resolve that here.

Phase 3 — Citation alignment. We fix the entity: one consistent description of what you sell and who it’s for, across your site, directories, review profiles and the comparison pages in your category. Then we fill the fit gaps — the “best for” pages, honest comparisons, alternatives pages and customer proof your buyers’ questions call for.

Phase 4 — Measurement. Every month we rerun the same prompt set per engine and track the five things that matter: recommended or not, shortlist position, fit accuracy, description accuracy, and who took your place. When something moves, we trace it back to the source that moved it.

The phases run in order because each one depends on the last. Content on top of an unresolved entity doesn’t compound. Measurement without a fixed baseline tells you nothing.

If you want a first look at what the models recommend when your buyers ask, start with a free Shortlist Check: three or four of your buyers’ questions in one engine, run before we talk. Book a call.

Quick answers

What is an AI recommendation? When an AI model names your product as an option in response to a buyer’s question, rather than only citing your content as a source.

Is being cited the same as being recommended? No. A citation means your page was used as a source. A recommendation means the model suggested your product. Many companies have the first without the second.

How do I get recommended by ChatGPT? Make your category clear, state who you’re built for, keep your description consistent across the web, publish bottom-of-funnel pages like comparisons and “best for” pages, and keep your pricing, integrations and proof current.

Can you rank in ChatGPT? Not in the traditional sense. ChatGPT writes answers instead of listing results. The equivalent goal is being recommended — and described correctly — when buyers ask the questions they ask before choosing.

Why does AI recommend my competitor and not me? Usually because they’re easier to categorize, described more consistently, have pages that answer fit questions, or appear in the comparison lists the model draws from.

How do I track AI recommendations? Run a set of unbranded buyer questions monthly across each engine. Log whether you’re recommended, your shortlist position, whether fit and description are right, and who appears instead.

Do SEO and AEO still matter for recommendations? Yes. SEO gets your pages retrieved. AEO makes sure the resulting answer recommends you and describes you correctly. See AEO vs SEO and AEO vs GEO.