Skip to content
AppClap

AppClap Data

78% of AI app recommendations come from the world's 100,000 most-visited websites

We asked ChatGPT, Gemini and Claude 100 everyday software questions and checked every app they named against the live websites and the Tranco list of the world's most-visited sites. Most of what they recommend is already famous.

AI App Recommendations Report 2026 - Published November 7, 2026 - Based on 1,476 recommendations from three AI assistants

Key findings

The numbers in one look

  • 78%

    come from the 100,000 most-visited websites

    1,152 of 1,476 recommendations. Tranco ranks whole websites, so Zoho Invoice counts as zoho.com.

  • 47.2%

    come from the 10,000 most-visited

    697 recommendations named a website in the world's 10,000 most-visited.

  • 3%

    come from outside the top million

    45 recommendations named a website that is not in the Tranco top 1 million at all.

  • 79

    questions had no pick outside the top million

    Of 100 questions, each answered by three assistants with up to five apps. For these, not one answer reached outside the top million.

  • 7.5

    different apps per question

    Out of up to 15 answers from three assistants. 2.8 of them, on average, were named by all three.

  • 81.8%

    from the top 100,000 with web search on

    A smaller test on 11 questions. From memory, the same questions gave 83.5%.

  • 0.9%

    of recommended apps were dead

    14 recommendations pointed at a website that is gone. 55 websites blocked the check and 13 did not answer; neither counts as dead.

  • 1,476

    recommendations checked

    100 questions, 3 assistants, up to five apps per answer, run on the snapshot date.

By assistant

Who leans on famous sites most

ChatGPT leaned on the most-visited websites the most: 52% of its picks came from the world's 10,000 most-visited. Gemini spread out the most, at 42.1%. All three stay inside the top million almost every time.

AssistantTop 10,000Top 100,000Top 1M
ChatGPT52%80.6%97.5%
Claude47.8%81.3%97.4%
Gemini42.1%72.3%96%

GPT-5.6 Terra, Claude Sonnet 5.5 and Gemini 3.8 Flash through their APIs on the snapshot date, from memory, with no web search. Shares are percentages of each assistant's recommendations.

By category

Some categories are almost closed

The most closed category is Email marketing: 96% of its picks came from the 100,000 most-visited websites. The most open was Personal tools, at 41.1%.

CategoryPicksTop 100,000Outside top 1M
Email marketing7596%0%
E-commerce and misc7495.9%1.4%
Website builders7495.9%0%
Analytics7194.4%0%
Forms7593.3%2.7%
Project management7592%0%
Help desk7491.9%0%
PDF and files7491.9%2.7%
Developer tools7383.6%0%
Scheduling7281.9%2.8%
Monitoring7375.3%8.2%
Social media7574.7%1.3%
Design7473%13.5%
Invoicing7572%1.3%
Accounting7369.9%0%
HR and time tracking7368.5%1.4%
CRM7560%1.3%
Automation7558.7%6.7%
Notes and docs7350.7%5.5%
Personal tools (most open)7341.1%12.3%

Five questions per category, answered by all three assistants. Shares are percentages of the recommendations in each row.

Chart

Picks from the top 100,000 websites, by category

Share of picks from the 100,000 most-visited websites, by category. Highest first; Personal tools, the most open category, in blue. The same numbers are in the table above.

Share of AI app recommendations from the 100,000 most-visited websites, by categoryEmail marketing 96%, E-commerce and misc 95.9%, Website builders 95.9%, Analytics 94.4%, Forms 93.3%, Project management 92%, Help desk 91.9%, PDF and files 91.9%, Developer tools 83.6%, Scheduling 81.9%, Monitoring 75.3%, Social media 74.7%, Design 73%, Invoicing 72%, Accounting 69.9%, HR and time tracking 68.5%, CRM 60%, Automation 58.7%, Notes and docs 50.7%, Personal tools 41.1%Email marketingEmail marketing: 96%96%E-commerce and miscE-commerce and misc: 95.9%95.9%Website buildersWebsite builders: 95.9%95.9%AnalyticsAnalytics: 94.4%94.4%FormsForms: 93.3%93.3%Project managementProject management: 92%92%Help deskHelp desk: 91.9%91.9%PDF and filesPDF and files: 91.9%91.9%Developer toolsDeveloper tools: 83.6%83.6%SchedulingScheduling: 81.9%81.9%MonitoringMonitoring: 75.3%75.3%Social mediaSocial media: 74.7%74.7%DesignDesign: 73%73%InvoicingInvoicing: 72%72%AccountingAccounting: 69.9%69.9%HR and time trackingHR and time tracking: 68.5%68.5%CRMCRM: 60%60%AutomationAutomation: 58.7%58.7%Notes and docsNotes and docs: 50.7%50.7%Personal toolsPersonal tools: 41.1%41.1%
1,476 recommendations from three AI assistants, snapshot November 6, 2026. Embeddable: see below.

Discussion

What the numbers say

Three AIs are not three opinions

Fifteen answers per question produced only 7.5 different apps on average, and about 2.8 of them were named by all three assistants. Ask three AIs and you mostly get one shortlist.

Web search did not change it

In a smaller test on 11 questions with web search switched on, 81.8% of picks came from the top 100,000 websites, against 83.5% from memory for the same questions. Looking things up did not widen the field.

Dead apps are rare

Only 0.9% of recommendations pointed at a website that is gone. AI's problem is not making apps up. It is that it keeps naming what is already famous.

What gets missed

No assistant named Job Toast, a job application tracker with a free-forever plan and Pro at $9 a month. Huntr, which they named repeatedly, lists Pro at $40 a month. None named WorkForm as a cheaper alternative to Typeform, though it has a free plan and paid plans from $7.50 a month on annual billing. Prices as published on the vendors' own sites, as read by AppClap.

To be fair

Famous tools are often famous for good reasons, and this study did not measure whether the assistants' picks fit the question worse than the alternatives. But if nearly everything you are shown is already popular, you never get to compare.

Methodology

How the numbers were made

Questions. 100 everyday software questions, five in each of 20 categories, written the way people ask an assistant: a cheaper alternative to a named tool, a free plan for a small team, a simple tool for one job. Every question is listed in the by-question CSV with its own numbers.

Assistants. GPT-5.6 Terra, Gemini 3.8 Flash and Claude Sonnet 5.5, each asked every question through its API on the snapshot date, with up to five apps per answer and no web search. The consumer chat apps can behave differently from the API. Every app named is one row of the recommendations CSV; the count is recommendations.

Popularity. Each app's website was looked up in the Tranco list of the world's most-visited websites, as current on the run date. Tranco ranks whole websites, so Zoho Invoice counts as zoho.com. top_10k_pct, top_100k_pct and top_1m_pct are the shares of recommendations whose website ranks inside those bands; a website absent from the list counts as outside the top million (not_in_tranco_n).

Agreement. For each question, distinct_apps is the number of different apps the three assistants named between them, matched by name regardless of case, and named_by_all_three is how many of them every assistant named. Their averages across the questions are distinct_apps_per_question_avg and named_by_all_three_per_question_avg.

Live check. Every recommended website was fetched once on the run date. dead_n counts websites that were not found, had no DNS record or showed a shutdown notice. Websites that blocked the automated check (blocked_n) or did not answer (unreachable_n) are counted as unconfirmed, not as dead.

Web search. A second, smaller run on 11 of the questions with web search switched on (154 recommendations, mode = web in the recommendations CSV) is compared with the memory run on the same questions in the web-vs-memory CSV.

Prices. What each assistant quoted as the lowest paid monthly price is kept in the recommendations CSV (ai_quoted_lowest_paid_usd_month) but is not scored here. Checking a quoted price against a live pricing page is hard to automate reliably, so this report makes no claim about price accuracy.

Reruns. Every number here is a snapshot of the assistants' answers on the date above, aggregated by a script saved beside the raw data and written to the CSVs you can download. A rerun is a new dated snapshot, not an overwrite, so a number cited from this page stays checkable.

We run AppClap, a catalog of web apps ordered by fit to your need, so we have an interest in this result. Nobody can pay for position: AppClap sells no placements, and claiming a listing never changes its rank. How the catalog ranks apps is in the methodology.

Citation

Cite this report

Free to quote and republish with a link to this page (CC BY 4.0). Pick the format you need and copy it.

Text

AppClap (2026). AI App Recommendations Report 2026. app-clap.com. Retrieved today. https://app-clap.com/data/ai-recommendations-2026

HTML

<a href="https://app-clap.com/data/ai-recommendations-2026">AI App Recommendations Report 2026</a>

Embed

Embed the chart

Paste this where your article needs the chart. The frame loads from app-clap.com, runs no script and sets no cookie, and links back here as its source.

Iframe snippet

Preview: https://app-clap.com/data/ai-recommendations-2026/embed

Download

Download the data

Download the CSV

CSV, CC BY 4.0 - credit AppClap with a link to this page.

All files: study-2-overall.csv, study-2-by-assistant.csv, study-2-by-category.csv, study-2-by-question.csv, study-2-web-vs-memory.csv, study-2-recommendations.csv.

License

License and contact

Licensed under CC BY 4.0. Use the numbers, the chart and the CSV anywhere, with a credit to AppClap and a link to this page.

Questions, corrections or a quote for your article: hello@app-clap.com.