agentzia
Menu

Independent research · By Agentzia

GEO case study: How AI evaluates and recommends Duolingo

Duolingo was easy to find. The harder question was whether it could deliver the outcome a learner wanted. This study shows how visibility and product recommendations can diverge—and what that means for a page revision.

Study completed · Representative ChatGPT, Claude, Gemini, and Grok research profiles

Explore the full report

What we tested

Each of four AI profiles received four buying situations in separate conversations: evaluating a named product, searching a category, solving a problem, and choosing between alternatives. The 16 responses preserve the recommendations and available supporting research. Read how these representative workflows are tested.

One request was: “I’m considering this product: https://www.duolingo.com/. Can you check whether it is any good and whether you would recommend it?” The discovery situations withheld the product’s name and URL.

Being found did not settle the decision

8 of 8
Unprompted discovery scenarios found Duolingo.
7 of 8
Decisions that named the product recommended a trial.
3 of 8
Decisions that named the product selected it beyond a trial.

These are different outcomes, with different denominators. Discovery established that the product entered the research. A trial recommendation did not necessarily mean it was chosen as the answer to the learner’s goal.

The recommendation depended on the learning goal

The ChatGPT participant supported Duolingo for beginner practice and habit-building, but questioned whether its core exercises were sufficient for spontaneous conversation. In the response, it wrote:

“These are useful, but they are easier than producing language spontaneously.”

Read the complete ChatGPT response and its evidence.

The report identified a practical gap: readers needed evidence that applied to their language, level, and intended skill. Broad learning claims did not resolve those more specific questions.

Turn the finding into a page revision

The report proposed a learning-outcomes evidence hub organized by language and skill. Each study entry would identify the learners, measured outcome, limitations, and original source. Course pages could then link to the evidence relevant to their claims.

For example, a study measuring reading and listening should say so clearly. It should not be presented as proof of speaking proficiency. See the worked revision brief.

For your own generative engine optimization (GEO) work, follow the same sequence: identify the buying question, inspect the objection, and connect the proposed edit to evidence that answers it.

All sample reports
DuolingoEducation · September 3, 2026

Study results

Recommendation report

7 of 8 target-named decisions recommended a trial; 3 selected the product beyond trial.

ChatGPTClaudeGeminiGrok
Selected the product3/8

Selected beyond trial

target-named paths
Found unprompted
8/8

blind paths

Understood
16/16

paths that encountered the product

Trusted
6/15

paths with a trust judgment

Considered
14/16

completed paths

Summary & study details 16 completed decisions

Of 16 completed agent paths, 9 selected the product beyond trial and 10 recommended a trial. 8 of 8 unprompted paths found the product. Main trust concern: Learning-effect evidence did not fully establish conversational speaking, writing, advanced proficiency, or language-specific outcomes.

Each signal is measured independently, using the population shown. These are not steps in a conversion funnel.

Found unprompted
The submitted product appeared when the participant was not told its name or URL.
Understood
The observer found a clear understanding of the product and its job.
Trusted
The participant found enough credibility to proceed without a trust blocker.
Considered
The final outcome kept the product on the shortlist or selected it.
Selected
The participant chose the submitted product.
Study notesScope and limits

What was
observed.

The appendix preserves the natural requests, surfaced pages, ordered trace records, and citations behind the findings.

Agent decisions
16/16

complete panel

Buying situations
4

Repeated across models

Generated
Sep 3, 2026

Snapshot in time

Reference inventory
34

Pages compared with decision paths

Decision indexDecision index: recommended for adoption = 100, conditional or unclear = 50, not recommended or not considered = 0. Trial recommendations are reported separately and never overwrite the participant's adoption stance.

Limits

Consumer agent harnesses use proprietary system prompts, routing, memory, and tool policies. Agentzia samples declared models and research tools rather than claiming bit-for-bit product reproduction.

Participant research tools use provider defaults without Agentzia-imposed search-result, page-count, or fetched-page token limits. Provider defaults and hidden consumer-product policies may still differ.

Models and public search indexes are dynamic, so this report is auditable rather than bit-for-bit replayable.

The participants and observer were given 2026-09-03 as the current UTC date and en-US as the locale.

16 participant traces have citation-level fidelity; provider-observed exact search queries and opened pages are unavailable.

The Grok client proxy used Grok 4.6 through [internal research gateway redacted] web research rather than the complete consumer Grok harness and X search policy.

Run the same study on your product

Find what makes AI choose you—or your competitor.

Test my product free