YouTube in AI answers: Perplexity dropped videos, Gemini now cites them most

In July 2026, youtube.com was the most cited source on both Gemini and Perplexity across our surveys. Two months later, the two engines have gone opposite ways. Between August 10 and 17, Perplexity stopped putting videos among its sources: on the same questions, asked word for word every week, the share of its sourced answers citing at least one video fell from 61.5% to 0.7%. Gemini, on the model Google now serves to the public, cites a video in 46.5% of its sourced answers, and youtube.com has become the domain it cites most. This study dates the break, traces it to its source, and says what it changes for a video meant to reach AI answers. It continues our July study of YouTube in AI answers and puts its findings back to the test.

What was measured

The corpus is the answer history of 19 of our own trackers, on two fields that have nothing in common: travel destinations, asked in English (9 trackers), and AI visibility, asked in English, French, Spanish, German, Italian and Brazilian Portuguese (10 trackers). Every question is asked word for word, through each engine's official API, on a fixed cadence, and every answer is archived with its cited sources, in order, dated. The window runs from July 10 to September 18, 2026, on every engine we survey.

The comparison sets the surveys up to August 10 against the surveys from August 17, and keeps, for each engine, only the trackers surveyed on both sides: a change in the mix of questions cannot pass for a change in the engine. The share below is the share of answers with sources that cite at least one YouTube video.

Two lines moved, in opposite directions. The others barely stirred: ChatGPT and Mistral cite no video on either side, Claude and Grok one now and then. Gemini's second column blends two models, and that detail is the whole story of its line, taken apart below.

Perplexity: one week in August

The break is dated to the week. On the nine travel trackers, 39 of Perplexity's 54 answers cited a video on August 10; on August 17, none of the 54 did. On the AI visibility trackers, the last regular video citations appear in the surveys of August 7 and 8, and from August 18 they are gone, two exceptions aside. Every field and every language switched in the same week:

In citations, youtube.com fell from 548 of 7,682, which made it the most cited domain, to 11 of 5,235. On the Atlas, which scores every source on a 0 to 100 scale across all the measurements Epovest runs over 30 days, its AI Authority on Perplexity went from 100 in July to 9 today.

The change happened on Perplexity's side. The request our surveys send and the model that answers it are the same as in June. The raw results Perplexity returns, which our archive keeps whole, carry the change themselves: the videos are missing from what the engine hands back, before any processing on our side. Perplexity's API changelog carries no entry on the subject between July and September.

Three details give the change its outline:

• User content stays. Reddit remains among the sources of 49% of answers (68% before) and LinkedIn of 14% (11% before). What left in the same week: YouTube videos, and with them Facebook groups and Instagram reels, from 61 answers to 1 and from 20 to none.
• Perplexity cites as much as before. An answer carries 19.0 sources on average, against 15.6 before: the room videos left went to other pages, vendor pages, software directories and trade press on the AI visibility questions.
• Two exceptions remain. Two answers out of 276 still cite videos, five and six at once, both to questions about ChatGPT itself, such as "How do I get recommended by ChatGPT?".

Gemini: the public model put YouTube first

Our surveys follow the model each lab serves to the public by default, and on August 22 Gemini's changed. The rise is tied to that switch, not to the calendar. On the 10 trackers surveyed with both models:

In the days just before the switch, from August 17 to 22, the previous model was citing a video in 7.1% of its answers. On the current model, youtube.com carries 4.3% of all citations and is the most cited domain, ahead of reddit.com. On the Atlas, its AI Authority on Gemini holds at 100.

What Gemini cites is specific. The 74 citations point to 22 videos from 19 channels:

• Comparisons first. 41 of the 74 citations go to a video whose title announces a list or a comparison of tools, such as "5 Best Generative Engine Optimization Tools 2026" or "7 Best AI Search Tracking and Visibility Tools". The two most cited videos carry 24 citations between them.
• Often in first position. In 22 answers, the video is the first source listed.
• Specialist channels. Next to Ahrefs, HubSpot Academy and Hootsuite, most channels are small specialist ones: Jay's AI SEO Tips, GEO Mastery with Adam, Surfside PPC.
• One language for six markets. 27 of the 28 citations on questions asked in French, Spanish, German, Italian or Portuguese went to videos in English. The most cited video answered questions in five languages.

The travel trackers, surveyed monthly, were last asked on August 17, before the switch: Gemini's figures above come from the AI visibility questions.

What held since July, and what changed

The July study drew four lessons from its corpus. Two months on:

• The title opens the door: it holds. The videos Gemini cites carry the question in their title, in the form the question takes. Asked "What are the best AEO tools for answer engine optimization?", Gemini cites "The Best Tool to Use for Answer Engine Optimization". The same measure holds for pages: the title shaped like the question wins.
• The spoken track is what gets read. In July, the evidence came from Perplexity's retained passages, timestamped transcripts. On Gemini, the route is Google's video indexing, which reads what a video says and detects its key moments.
• Two engines quoted video: one does today. Perplexity closed its door in a week, without an announcement.
• Language follows the market: now nuanced. In July, French questions surfaced French videos first. On today's Gemini, for AI visibility questions, an English video answers six markets.

What this makes worth doing

• Aim the video at Gemini, and keep measuring the others. Perplexity closed its door in a week, and a door can reopen as fast. Your own series, engine by engine, tells you where a video works today.
• Title it as the answer to the choice. When the market asks which tool, which destination or which provider, the video that enters Gemini's sources is the one that compares: "N best ... for ...", in the words of the question.
• Be in the comparisons others make. The videos Gemini cites are comparisons made by specialist channels. A brand named in them travels into the answer with them, and one left out stays out, which is the ordinary way a competitor lands in the answer instead of you.
• Start in English when the field speaks English. On AI visibility questions, one English video answered six markets on Gemini. A market with videos in its own language still deserves its own.
• Date the move, then read the effect. A published video is a move worth recording in your Logbook. The next surveys read against it, and the Atlas shows whether the sources behind your questions moved.

In July, a video platform was the most cited source of Gemini and Perplexity alike. In September, it is the most cited source of one and nearly absent from the other, with no announcement in between. The weight of a source in AI answers belongs to one engine at one date, and our method is built on that: ask the market's questions on a fixed cadence, archive the answers with their sources, and let the series say where to act.

Sources

• Perplexity: API changelog
• Google AI for Developers: Grounding with Google Search
• Google Search Central: video indexing and key moments