On this page
Ask Claude “what should a small SaaS use for uptime monitoring?” and it will name three or four products with a short reason for each. If yours is missing, a buyer who trusts Claude just got a shortlist without you on it. Most of the AEO advice floating around was written with ChatGPT in mind. Claude behaves differently in a few ways that matter, and if you understand those differences you can stop guessing and work on the things that actually change its answers.
How Claude decides which products to name
Claude is a large language model. When you ask it a recommendation question, two very different things can happen depending on where you are asking from.
Without web search, Claude answers from parametric knowledge: everything it absorbed during training, filtered through how it was tuned to respond. In this mode it tends to name products it has seen described many times, in consistent terms, across sources it treats as reliable: documentation, well-known review platforms, established publications, developer forums, GitHub. It is also noticeably cautious. If Claude is not confident a product exists or does what you say, it would rather leave it out or hedge (“there are also newer tools in this space”) than risk inventing details.
With web search enabled, which is the default experience for many users on the Claude apps, Claude retrieves live pages and grounds its answer in them, showing the sources it used. Now the question is not “has Claude heard of you” but “are you on the pages Claude just pulled for this query.” That behaves much closer to Perplexity, and the tactics from how to rank in Perplexity answers transfer almost directly.
Practically, this means you are running two campaigns at once:
- The slow campaign: become a consistently described entity on the web so that the next model version knows you.
- The fast campaign: get onto the specific third-party pages Claude retrieves for your buyer questions today.
Everything below serves one or both.
Step 1: Find out what Claude says right now
Do not start with content. Start with a baseline, otherwise you will not know whether anything you do later worked.
Write down 20 to 50 prompts your buyers would plausibly ask. Mix the formats:
- Category questions: “best [category] for [audience]”
- Alternatives: “alternatives to [market leader]”
- Comparisons: “[competitor A] vs [competitor B]”
- Problem-first: “how do I [the job your product does] without hiring someone”
Run each one in Claude two ways: in a fresh chat with web search on, and with it off (or via the API without tools). For every run, note three things: were you named at all, what position you appeared in, and how Claude described you. That last one matters more than founders expect. Being named as “a lightweight option for teams under 10 people” when you sell to enterprise is a positioning problem, not a visibility win.
Then run the same prompts a second and third time. Answers vary between runs, which is why multi-sampling is standard practice for anyone measuring this seriously. One appearance in one run is noise. Being named in 7 of 10 runs is a signal.
Step 2: Make your product a clearly defined entity
Claude’s caution works against unclear brands. If the web describes you three different ways, or barely describes you at all, Claude either skips you or fills the gap with something plausible and wrong. Your job is to make the answer to “what is [product]?” identical everywhere Claude might look.
Concretely:
- One canonical definition. A single sentence in the form “[Product] is a [category] for [audience] that [core job].” Put it in your homepage hero, your about page, your meta description, and your social bios. Do not get clever with it. Claude has to be able to repeat it.
- Structured data. Add Organization JSON-LD with the same name, description, URL, and
sameAslinks to your profiles. It costs an hour and removes ambiguity about who you are. - Third-party profiles that agree with you. G2, Capterra, Crunchbase, LinkedIn, Product Hunt, GitHub if you have any open source. Same name, same one-line description, same category. These are exactly the kinds of pages models are trained on and retrieve from.
- A Wikidata entry if you can justify one. It is a small thing, but it anchors your product as a named entity with a category and a website, which is precisely what entity extraction inside these systems is looking for.
None of this is glamorous. It is the difference between Claude confidently saying “X is a Postgres backup tool for solo developers” and Claude saying “there are also some smaller tools in this space.”
Step 3: Get onto the pages Claude retrieves
For the fast campaign, the target list is finite. Across your prompt set, the same domains keep showing up as sources: a couple of niche blogs, one or two review platforms, a comparison post that ranks well, maybe a Reddit or Hacker News thread. Those pages decide the search-grounded answer.
Work them in this order:
- Update pages you control. If you have a “[Product] vs [Competitor]” or “best [category] tools” page, make sure it is genuinely useful, current, and honest about where you fit. Claude weighs pages that read as balanced. A page that says “we are the best at everything” is less likely to be leaned on than one that says “we are best for X, competitor is better for Y.” Publishing comparison and best-of pages is one of the highest-return moves for a small team.
- Get listed where the shortlists live. Reach out to the authors of the listicles that keep appearing as sources. Offer a real reason to include you: a specific differentiator, a free account to test, a use case their existing list does not cover. Most of these pages are refreshed once or twice a year, and the person maintaining them usually wants the list to be complete.
- Show up in the conversations. Reddit threads, indie hacker communities, Stack Overflow, GitHub discussions. Not with drive-by promotion, but with actual answers to actual questions where your product is one relevant option. These threads get retrieved and, over time, trained on.
- Earn a few press or newsletter mentions in your niche. You do not need TechCrunch. You need the three newsletters your buyers actually read.
Perplexity’s citation lists are the cheapest way to build this target list, because it always shows sources. Claude with search shows them too, so use both.
Step 4: Make your own site easy for Claude to use
When Claude does retrieve your pages, it needs to be able to read them and pull out facts. That is where the technical basics matter:
- Allow AI crawlers in robots.txt. Anthropic’s crawlers, along with OpenAI’s and Perplexity’s, are commonly blocked by default in some hosting and firewall setups. Check yours. Blocking them means your own site cannot be a source, and you are left relying entirely on what third parties say about you. The concept of AI crawlability is exactly this.
- Publish specific, citable facts. Numbers, limits, pricing, integrations, what is on the free plan. “Fast and reliable” gives Claude nothing to say. “Backups every 15 minutes, restore in under 5, free up to 3 databases” gives it a sentence it can repeat and attribute.
- Write “what is X” definitions in plain prose for your category and your product. Claude generates recommendation answers by pattern-matching to definitions like these.
- Use semantic HTML. Real headings, real lists, real tables. Content hidden behind heavy JavaScript or rendered as images is invisible to a text-first reader.
- Consider an llms.txt file. It is not a ranking factor and no vendor has promised to honor it, but it is a low-cost, machine-readable summary of what your site is and where the important pages are. Nothing to lose.
What Claude does differently from ChatGPT and Perplexity
A quick comparison, because founders keep asking whether they need three separate playbooks:
| Claude | ChatGPT | Perplexity | |
|---|---|---|---|
| Default grounding | Training knowledge, web search when enabled | Training knowledge, web search when enabled | Live retrieval, always |
| Shows sources | Yes, when it searched | Yes, when it searched | Always, numbered |
| Willingness to name unfamiliar products | Low, prefers to hedge | Moderate | High if a retrieved page mentions you |
| Fastest lever | Presence on retrieved third-party pages | Presence on retrieved pages plus training-time mentions | Presence on retrieved pages |
| Slow lever | Consistent entity signal across the web | Same | Less relevant |
The foundations are shared. The main Claude-specific adjustment is that consistency and verifiability buy you more, because Claude is quicker than the others to leave out a product it cannot vouch for. If you get the entity work right, ChatGPT and Perplexity benefit too. Only the measurement needs to be per-engine.
Track it, or you will not know
Everything above takes weeks to months to show up in answers, and Claude’s answers move between runs and between model versions. If you check once in a browser and then check again in two months, you will not be able to tell whether your work did anything or a model update happened to shuffle the deck.
The practical setup is boring and effective: run your prompt set against Claude (and the other engines) on a schedule, sample each prompt several times, and log mention, position, and description. Compare week over week. That is what prompt tracking means in practice, and it is the difference between AEO as a hobby and AEO as something you can report to yourself in a spreadsheet.
You can do this by hand for a while with a text file and a Sunday morning. When that stops scaling, AskAiRank runs the same prompts across Claude, ChatGPT, Perplexity, and Gemini automatically, samples each one, and shows you the trend line and where competitors are being named instead of you.
Next step
Pick 20 buyer prompts today, run them in Claude with search on and off, and write down what you find. Then open the sources Claude cited on the prompts where you were missing and list the pages you are not on. That list is your plan for the next month: fix your entity definition, get onto those pages, and check again in four weeks against the same prompts.
Frequently asked questions
Answers about what makes Claude recommend one SaaS product over another.
Sometimes. On the Claude apps, web search can be enabled and Claude will retrieve live pages when the question benefits from it, and it shows the sources it used. Through the API, without a search tool wired in, Claude answers from what it learned in training. So you have to work both channels: earn a place in the training-time signal, and be present on the pages Claude retrieves when it does search.
For the search-grounded path, weeks: as soon as your product appears on the comparison and review pages Claude retrieves for a question, it can be named. For the training-knowledge path, months to a year, because it depends on the next model version being trained on a web where your product is mentioned consistently. Most founders see search-grounded mentions long before parametric ones.
Usually because the web has conflicting or thin descriptions of you. Claude fills gaps with the most plausible pattern, which can mean borrowing facts from a similar-sounding product. Fix the source: a clear one-sentence definition on your homepage, matching descriptions on G2, Crunchbase, LinkedIn, and Wikidata, and Organization JSON-LD on your site.
The foundations overlap heavily: clear entity definition, presence on trusted third-party pages, and crawlable content help everywhere. Claude is more conservative about naming products it is unsure of, so consistency and verifiability matter a bit more. Track them separately, but you rarely need separate content programs.
Run the same 20 to 50 buyer prompts against Claude on a schedule and record whether you are named, in what position, and how you are described. A single manual check tells you almost nothing because answers vary between runs. A weekly trend line across the full prompt set is the only signal worth acting on.
