How to Get Your App Recommended by ChatGPT, Gemini, and Claude

A meaningful slice of app discovery no longer starts in the App Store search box. It starts with someone typing "what's a good app to restore old family photos?" into ChatGPT, Gemini, or Claude — and downloading whatever the assistant names. If the assistant has never read a clear sentence about your app, you are not in that conversation.

This isn't classic ASO, and it isn't classic SEO either. An AI assistant doesn't run your keywords through a ranking algorithm; it synthesises an answer from whatever it can read about your app across the web — your store listing, your website, reviews, forum threads, and articles. The discipline of making that material machine-legible has a name now (answer-engine optimisation, or generative engine optimisation), but the practice is mostly about clarity. Here's the playbook I apply to my own apps, including Heirloom.

How AI assistants actually pick apps

When someone asks an assistant for an app recommendation, three things happen: the model interprets the intent (the job to be done, not the keyword), it retrieves whatever indexed text best describes apps that do that job, and it synthesises a short answer naming one to five apps with reasons. AppTweak's analysis of AI app recommendations found that App Store product pages are the single most-cited source — roughly 38% of citations, with Google Play adding about 9.5% — followed by the open web: review sites, blogs, forums, and developers' own pages.

Two consequences follow. First, your store listing is now read by machines that quote it, not just skimmed by humans. Second, the rest of the web's sentences about your app matter as much as the listing itself.

Write your store listing like an answer, not an ad

Models extract three things from a listing: who the app is for, what it does, and when to recommend it. Make those impossible to miss. The pattern that works is almost embarrassingly plain:

Heirloom is an AI photo restoration app that helps families repair old, damaged, or faded photos by restoring, sharpening, and colourising them on iPhone.

Persuasive copy ("Rediscover your memories like never before!") gives a model nothing to extract. Concrete nouns — scratch repair, colourisation, memorial portraits, offline puzzle, no ads — are what get quoted. Keep the first paragraph of your description factual and specific; poetry can come after.

Give every app a real page on your own site

Store pages are rented; your site is owned. Every app should have a canonical page that states the same facts in crawlable HTML, carries SoftwareApplication structured data, links to both stores, and answers the obvious questions in literal FAQ form. That's how every app page on this site is built — the Heirloom page, for instance, ships SoftwareApplication and FAQPage JSON-LD generated from the same data file that renders the page, so the markup can never drift from the content.

Why FAQs specifically: assistants love lifting question-and-answer pairs, because the extraction work is already done. A page that asks "Does it work offline?" and answers in one sentence is a page an AI can quote verbatim and attribute.

Structure your site for machines

The checklist I apply, in rough order of value per hour:

1. JSON-LD everywhere it's true — Organization and WebSite on the root, SoftwareApplication per app, BlogPosting and FAQPage per article. Generate it from data, don't hand-write it twice.
2. Question-shaped headings with first-sentence answers. An H2 that asks the question a user would ask, answered directly in the next sentence, is simultaneously good writing and good extraction.
3. An llms.txt file — a plain-text map of what your site and products are, served at /llms.txt (the convention proposed at llmstxt.org). It costs an hour and gives crawlers a clean, unambiguous summary. This site serves one, generated from the same data files as the pages.
4. Don't block AI crawlers in robots.txt unless you've genuinely decided that trade-off; being unreadable is the one guaranteed way to never be recommended.
5. Keep facts consistent across store listing, website, and socials. Contradictory descriptions make a model hedge — and hedging models name your competitor instead.

Earn sentences you don't write

Assistants weight independent sources — Reddit threads, niche blogs, comparison articles, review sites — precisely because developers don't control them. You can't fake this (and shouldn't: fabricated reviews and astroturfed threads are both store-policy violations and increasingly detectable), but you can earn it: ship something genuinely distinctive enough to mention, answer questions in communities where your users already are, and make it trivial for bloggers to describe your app accurately by publishing a clear press/about page. Specific, recent store reviews matter too — a review that says "repaired a torn 1960s wedding photo" teaches a model exactly when to recommend you.

What doesn't work

Keyword-stuffed descriptions, self-published "Top 10 apps" listicles that rank yourself first, cloaked pages showing different text to crawlers, and review incentivisation. All of these are either ignored by models, penalised by stores, or — worst case — quoted verbatim in ways that embarrass you.

The honest summary: AI recommendation is an amplifier for clarity. The apps that win are the ones about which the internet contains many short, true, consistent sentences. Write those sentences yourself where you can, and earn them where you can't.