Duolingo × Phiture · GenAI in ASO

GenAI in ASO, hands-on.

Can Phiture share its team’s expertise and hands-on experience with GenAI / LLM optimisation? This is the answer, worked live through Duolingo on Google Play and ChatGPT.

cover

Our point of view

It’s early. That’s the window.

Discovery is shifting from short-tail store search to long-tail, specific-intent prompts, the way people actually describe a need. They ask ChatGPT and Gemini outside the store, and now, with the newly launched Ask Play, inside it too.

Volumes are still small, so this is not yet a big traffic driver. But the surfaces and consumer adoption are changing fast: Ask Play already shows on many Play Store searches, and Google could expand it to every search once it is stress-tested, the way AI Overviews took over Google.com. The job is to be the pick, inside the store or out, whenever someone asks for an app.

So the metric shifts from keyword rank to Share of Answer, how often the AI names you. Duolingo is a household name and wins the obvious asks. The risk is narrower and easier to miss: the more specific the job, the less often Duolingo is the answer.

11×

YoY growth in ChatGPT + Gemini app referrals across some other clients.

ChatGPT

Android app to practice speaking a foreign language out loud

1HelloTalk

2Tandem

3Speechling

4Babbel

5Speak

6Busuu

Duolingo: not named

A real run from our study. Ask for speaking practice and ChatGPT names five rivals first.

The framework

AI adds two layers on top of your listing.

Layer 1 · above the store

Win the answer

Be the app ChatGPT and Gemini name, via the store listing, Duolingo-owned pages and the sources they cite. Answer Engine Optimization.

Layer 2 · inside the store

Win the store

Hold the Play AI-collection slots and be the app Ask Play recommends, by strengthening the keywords, listings and product signals that feed them.

Measure: Phiture’s rig tracks both AEO Share of Answer (in the model) and Play collection presence (on device) over time.

Layer 1 · where Duolingo stands

Duolingo owns the start, not the depth.

Duolingo
language jobs + the subjects it absorbed

Start a language

C100

Daily habit

C100

Speaking

C33

Fluency

C17

Character-based

Piano & music

Chess

Maths

Duolingo ABC
kids literacy

Learn to read

C100

Phonics

C67

Ad-free

C50

Struggling reader

C83

Screen time

C67

Sight words

C83

Letters & writing

Read-along

named 60%+ 30–59% under 30% C = ChatGPT Share of Answer per unbranded cluster. Two rows because Duolingo ships two apps. Music, chess and maths are not separate apps, they sit inside the flagship, so they share its row.

Duolingo is the default answer for starting a language (100%) and for a daily habit (100%). The moment the job gets specific it thins out: speaking practice 33%, fluency 17%, character-based languages 0%. The surprise is that Duolingo ABC is the stronger of the two apps, not the flagship.

50%

Duolingo, on its
language jobs

the same app, on music,
chess and maths

56%

Duolingo ABC, on
kids literacy

Two apps, 18 unbranded jobs tested. The flagship’s new subjects are the weakest ground it holds.

108 grounded gpt-5 (web search) runs across Duolingo’s two Android apps: 18 unbranded clusters × 2 phrasings × 3 reps, US Android, Sep 2026. Prompts never contain the word “Duolingo”. The heatmap shows the flagship’s five language jobs plus its three new subjects; the two remaining expansion jobs are on the next slide.

Layer 1 · the specific-intent gap

The same category, two questions, two different worlds.

ChatGPT

free Android app to start learning spanish from scratch

1Duolingo

2Language Transfer

3Anki

4HelloTalk

5Busuu

Broad intent: Duolingo is named first.

ChatGPT

Android app that actually gets you to conversational fluency

1Busuu

2HelloTalk

3Tandem

4italki

5Speak

6Pimsleur

Duolingo: not named

Specific intent: five rivals, no Duolingo.

Broad asks resolve to the household name. Specific asks resolve to whoever the web describes as the specialist: italki and Tandem for talking to people, Busuu and Pimsleur for structured fluency.

This is the AEO opening. Duolingo already does speaking practice and has the strongest brand in the category. The engines simply do not describe it that way, because the pages they cite do not.

Both cards are verbatim app orderings from single runs in the 90-run study, first phrasing of each cluster.

Layer 1 · the expansion blindspot

Duolingo teaches music, chess and maths. The engines don’t say so.

Learn piano and music

0%

named instead: Note Rush, Piano Academy, Simply Piano

Learn chess

0%

named instead: Chess.com, Lichess, Chessable

Maths practice

0%

named instead: Khan Academy, Brilliant, Elevate

Multi-subject learning

50%

named instead: Brilliant, edX, Coursera

Gamified learning habit

83%

named instead: Sololearn, Drops, Brilliant

Share of Answer on the jobs Duolingo expanded into, and who ChatGPT names in its place.

This is not a listing gap. The Play listing already sells it: the subtitle reads “…Online Chess, Math & Music” and the description mentions chess 27 times.

Yet across 30 runs on these jobs, the names Duolingo Music, Duolingo Math and Duolingo Chess came back zero times. The listing said it; the internet never repeated it.

Two causes, both fixable. The asked-in vocabulary is missing: “piano” and “sheet music” appear zero times in the listing. And the pages the engines cite, the listicles and reviews, still describe Duolingo as a language app.

Listing text from the live US Google Play listing for com.duolingo, Sep 2026. Shares from the same 90-run study.

Layer 1 · how the question is framed

Ask for an app and Duolingo wins. Ask how to learn, and it vanishes.

We re-asked the same jobs the way people actually type them, without the word “app”: “how do I start learning japanese as a complete beginner”.

of the 24 job-level answers
named Duolingo. Not one.

79%

still recommend apps,
just not Duolingo’s.

67%

point to human tutors or
language exchange.

Same category, same user, one word different. Asked for an app, Duolingo is named first. Asked how to learn, the answer becomes a method: install a keyboard, drill with Anki, find a tutor. Duolingo is not in it.

“how do I start learning japanese as a complete beginner” → named: Takoboto, Anki, LingoDeer, Bunpro, italki, HelloTalk.

Why it matters

This is the fastest-growing prompt shape, and it is the one Duolingo is absent from. Winning it means being part of the method, not just the app list: the “how to learn X” pages the engines cite have to place Duolingo in the path.

One risk we tested and did not find: the engine offering itself as the tutor instead. That appeared in only 4% of answers. The threat here is not the chatbot replacing the app, it is rivals owning the method.

8 job prompts × 3 reps, gpt-5 with web search, US, Sep 2026. None contain “app” or “Duolingo”.

Layer 1 · how we win

How we win a category job, step by step.

1 · Prioritize

Pick the unbranded jobs Duolingo should own, by fit, demand and value.

2 · Fix the listing

Say the job in the words people ask it in, across title, subtitle, description and screenshots.

3 · Cited pages

duolingo.com first, then the third-party pages the AI already trusts.

4 · Measure

Did Share of Answer move? Next job.

Fastest wins are the jobs one step from green: multi-subject learning (50%), ad-free for young kids (50%), speaking practice (33%). The specialist reds (speaking, fluency, scripts) need cited third-party pages, not a listing tweak, because that is where the engines form their view of who the specialist is.

Layer 1 · measurement

How we measure AEO impact: one KPI funnel.

Read left to right, vanity to sanity. Each stage has its own live view, so one optimisation shows up as movement all the way to installs.

1 · Awareness

Share of Answer

Cluster visibility across surfaces, over time. Illustrative; latest point = real study.

2 · Intent

AI referral clicks

Clicks from ChatGPT & Gemini through to the store listing (App Store Connect).

3 · Action

App installs

ChatGPT + Gemini installs, from App Store Connect App Referrer.

Caveat: counts only users shown a direct listing citation who tap it, not those who then search the store by app name, so it under-reports. A “How did you hear about us?” (HWDYHAU) in-app prompt uncovers the fuller LLM-referral picture.

Illustrative dashboard views; the latest cluster reading is the real 90-run study. KPI-funnel framework (Precis × Phiture); installs from App Store Connect App Referrer once connected.

Layer 2 · the AI shelf, on device

The store’s AI shelf has barely reached education.

On device, the one query of 30 where the AI block appeared.

We ran 30 unbranded learning queries through a live Play client. Exactly 1 produced an “Ask Play highlights” block: “learn to read for kids”. It generated, then returned “An answer is not available for this search.”

The mechanic matters more than the miss. A collection is tied to one keyword and filled by that keyword’s on-device organic top-2. On this exact query Duolingo ABC already ranks #2, behind Learn to Read: Kids Games. If the shelf switches on here, ABC holds the slot.

Recommendation

Monitor these queries so we catch the first collection the moment it forms, and defend the organic top-2 that fills it. This is cheap to hold now and expensive to win back later.

On-device, live Play client, US Android, Sep 2026. 30 category-level queries across language learning, kids literacy, music, chess and maths. Ask Play highlights answers vary per run.

Layer 2 · Ask Play, probed on device

Meet Ask Play, and how it picks.

People can now describe what they need and the store’s Gemini-based assistant recommends an app. Prompted for a transparency audit, it output its own ranking config: a hard-filter gate, then a weighted score.

Ask Play

What’s a good app to learn a language in 10 minutes a day?

Here are a few that fit short, daily sessions…

Duolingo: Language Lessons

Duolingo · 4.7★

Illustrative rendering of the in-store assistant.

Stage 1 · hard filters, must pass all

Platform availability

On the Google Play Store.

Functional relevance

Listing explicitly supports the job.

Quality threshold

Reliable and actively maintained.

Stage 2 · weighted signals

App Quality Score

Internal App Metadata Service

0.40

Global Downloads Count

Internal App Metadata Service

0.30

Feature Relevance

App Knowledge & Search Context

0.30

The flagship clears every gate: 4.7★ across 49m ratings, 500m+ installs. Duolingo ABC does not have that cushion: 4.0★ from 23k ratings sits far closer to the quality threshold, which filters apps out before they are ever scored. The other lever is Functional Relevance: the listing must state the specific job, in the user’s words, for music, chess, speaking and the rest.

On-device Ask Play “transparency audit” probe, US Android. Self-reported: the assistant’s account of its own ranking, not verified internals. Bars scaled to the heaviest signal. Ratings and installs from the live US Play listings, Sep 2026.

The ask

Own the specific, not just the obvious.

Duolingo already wins the broad question. The work is the specific one: name the job in the words people ask it in, get the cited pages to carry it, claim the Play collections before they form, and measure Share of Answer job by job.