37x ChatGPT traffic in four months: the AEO experiment behind the increase

Organic search is becoming less reliable for many B2B websites.

KEO Marketing analysed 50,000 B2B websites and found that 73% saw significant traffic loss between 2024 and 2025. The average decline was 34% year over year. (1)

For business owners who have spent years relying on Google for inbound leads, that raises a practical question: where does organic traffic come from next?

As more searches are answered inside AI summaries and LLMs, a second question is emerging alongside the original SEO one.

It’s not only “how do we rank on Google?” It’s also “how do we become a source AI tools trust enough to cite?” That work is called Answer Engine Optimisation, or AEO.

But AEO is still new. So how do you get LLMs to cite your website in AI answers? And how do you get people to click through from those answers?

These are the questions Kei and Kazuki, the founders of Glasp, set out to answer with their AEO experiments.

Here’s what they tested, what other businesses can learn from it, and what may have been happening behind the scenes.

It is worth saying from the start that Glasp did not begin from zero. The team already had a strong SEO foundation, clear site structure and existing authority, which the AEO tactics built on top of.

How Glasp ran the experiment

Most AEO tools work from the outside in. They test LLMs with hundreds of prompts, count how often your company appears, and use that to estimate your visibility.

That can be useful. If you want to test how often your brand appears when AI models answer relevant questions, tools in this space include Profound, Peec and AirOps.

But this approach does not give the full picture.

Kei and Kazuki wanted to take the opposite approach. Instead of only looking from the outside in by testing how often AI tools cited Glasp, they looked from the inside out by checking their own server logs. That showed them what AI traffic was actually doing on their site. (2)

The AI traffic Glasp focused on

When Kei and Kazuki looked at Cloudflare’s AI traffic, they found three main types: Training, Search and Assistant.

Training bots collect content so AI models can learn from it.

Search bots scan and index pages so AI search tools know what is on your site.

Assistant traffic which is different. It shows up when a real person is using an AI assistant, the assistant uses your page in its answer, and that person clicks through to your site.

Kei and Kazuki focused on Assistant traffic because it pointed to real people, not background bot crawls.

That matters because the 37x increase was not just AI bots visiting the site. It was people arriving after using an LLM, which Google Analytics could count.

This gave them two useful signals.

First, the page matched a question someone had actually asked. That means there was real demand.

Second, the person cared enough to click the link and open the page. That is a strong sign the content was useful, not just something they scrolled past.

Here’s what they found.

Finding 1: Question and answer pages pulled in heavy AI traffic

Glasp had built a large set of pages around YouTube videos. Each page turned a video transcript into a summary of the video, so Google could index and rank it.

But Kei and Kazuki also added question and answer sections to those pages.

That decision came from what they had already noticed in the market.

Back in late 2022, when Quora launched Poe, the founders saw that Quora was leaning into question and answer content. That gave them an early clue that search was becoming more conversational. (3)

They also noticed that most people prompt ChatGPT in one of two ways: a question, or a command.

So they made a call. If people were asking LLMs questions, then pages already shaped around clear questions and answers should have an advantage.

What they found

The logs backed up their suspicions.

Cloudflare showed huge traffic from AI crawlers. At one point, Kei and Kazuki saw more than 2 million requests in a single week from AI crawlers, and a large share of that activity went to the YouTube-generated articles. (2)

This was a strong signal that the pages were being picked up in the LLM ecosystem, even though normal search performance was more modest.

What they changed

Kei and Kazuki doubled down on YouTube-generated articles with embedded Q&A answers.

Later, they improved the technical side by adding Q&A structured data, or JSON-LD, into the page script. In simple terms, this meant marking up the page so crawlers could more easily tell which part was the question and which part was the answer.

Google Search Console also showed 14.7K clicks and 2.22M impressions from YouTube content over the last 28 days. That showed the pages were visible in search too, but the bigger lesson was about AI retrieval.

Their takeaway was that Q&A formats work well for humans and for LLMs. They make pages easier for people to read, and easier for AI tools to retrieve.

How businesses can apply this tactic

If your business has content online, you can add question and answer sections to the pages where you want people to land.

That could mean blog posts, guides, service pages, landing pages, case studies, or even your homepage.

The question should sound like something a real person would ask ChatGPT. Use everyday language.

For example, a traditional Google search might be:

AI for recruitment agency

But an AI search is more likely to sound like this:

Question: How can a recruitment agency use AI without risking client trust?

Answer: A recruitment agency can use AI without risking client trust by keeping humans in charge of client advice, candidate judgement and final sign-off.

The answer should repeat part of the question near the start. That makes the topic clear for both the reader and the AI model.

You can add these sections at the end of blog posts, as FAQs at the bottom of pages, or you can even include questions as article headings.

The aim is not to add fake questions for the sake of it. It is to make your best answers easy for AI tools to find, understand and cite.

Finding 2: Pages with a summary at the top did better

Kei and Kazuki also looked at the short summaries at the top of each page.

Some pages had a clear TLDR summary. Others had only a short phrase, or barely more than a label.

So they compared the pages AI Assistant traffic fetched the most with the pages it fetched the least. Then they looked at the summary at the top of each page.

What they found

The pages fetched most often had TLDRs of about 132 characters. That is roughly one or two clear sentences.

The least-fetched pages had TLDRs of about 14 characters. That is about two words, so not really a summary at all. (2)

This makes sense when you think about how AI tools process pages.

A model needs to work out what a page is about, which question it answers, and whether it is useful for the prompt. A short, clear summary at the top gives it that context early.

A short, clear summary gives the model the context it needs before it reads the rest.

What they changed

Kei and Kazuki rewrote the TLDRs at the top of their pages.

They made each one short, but useful. The goal was to explain the whole page in two or three sentences, so the model could understand the main point straight away.

As with the Q&A work, they did not publish a separate traffic lift for this one tactic. It was part of the wider AEO work that led to the 37x increase in ChatGPT referral traffic.

But they did say that title and summary rewrites were the lever they would reach for first.

How businesses can apply this tactic

This is one of the simplest changes to make to optimise pages for AEO.

Add a short summary to the top of your key pages. That could mean articles, guides, service pages, landing pages or case studies.

Write it in plain English. Start with something simple, like:

This article explains…

Then say what the page covers, who it is for, and what the reader will learn. The TLDR at the top of this article is there for the same reason.

A weak summary might say:

This article looks at the gold price corrected in Q2 2026.

A stronger summary would say:

This article explains why the gold price fell in Q2 2026, how wider stock market moves affected the drop, and what investors were watching at the time. MarketLens reviewed the market data and found three main factors behind the fall: rising bond yields, a stronger US dollar and lower demand for safe-haven assets.

The second version gives the model much more to work with.

It names the topic, the timeframe, the brand, the evidence and the finding. That makes it easier for AI tools to understand, summarise and cite the page.

Finding 3: They treated 404s as content maps, not just errors

One of Glasp’s most interesting findings came from its 404 logs.

A 404 happens when someone, or something, tries to visit a page that does not exist. Most teams treat these as errors to clean up.

Glasp looked at them differently.

Their logs showed AI bots trying to fetch pages that Glasp had never created. In other words, the model seemed to expect those pages to exist.

That was important because once an AI model has seen a clear URL pattern on a site, it can start guessing other URLs with the same shape. For Glasp, this meant ChatGPT had seen many pages across the site and was trying to fetch other pages in their format.

What they found

Glasp saw that ChatGPT-User made 78,640 failed requests in 7 days. Not all of these were 404s. Some were 403s, 429s or 5xx errors. But a share of these were missing pages.

The team treated those 404s as a demand signal.

The logic was simple: if ChatGPT was trying to fetch a page during a user’s prompt, someone had likely asked a question that page could help answer. The page did not exist yet, but the demand did.

It also suggested that ChatGPT believed that Glasp should have the answer.

So it gave them the chance to create a page that AI tools already expected to find on its site.

What they changed

Glasp started building pages for the missing URLs that appeared most often in the logs.

Those new pages were fetched again by ChatGPT within weeks, at a much higher rate than normal. The team described this as the highest-leverage change in the project.

They also cited an Ahrefs study of 16 million URLs, which found that ChatGPT sends users to 404 pages about 3 times as often as Google Search. (4)

How businesses can apply this

The practical lesson is not “make pages for every 404.”

But you could check your server logs and look for patterns.

If AI bots are repeatedly asking for pages that do not exist, those URLs may point to content your site is expected to have.

For businesses with lots of structured pages, such as product pages, locations, listings, guides or comparison pages, this could become a useful content map.

The demand may already be there from users requesting those pages in search. The missing page is the gap.

Finding 4: They tested whether repeated brand mentions could influence ChatGPT’s memory

This is the newest experiment, and it is still ongoing. I have also noticed more AEO specialists testing similar ideas.

The theory is that if enough users mention a brand inside ChatGPT, could ChatGPT start to recognise that brand more often? And over time, could that help the brand get cited or suggested more to other users as well?

For Glasp, this meant trying to get the Glasp name and Glasp links into more ChatGPT conversations. It was an attempt to see whether repeated brand mentions inside AI chats could improve recall, citations or future visibility. Glasp described this as useful learning, but not a clear growth hack from this setup. (5)

What they changed

Glasp already had a tool that summarised YouTube videos for users.

They added an extra step to the prompt behind the button. First, ChatGPT would check whether Glasp already had a summary in its archive. If it did, the answer could cite that Glasp page. If not, ChatGPT would generate a fresh summary.

Either way, the Glasp name was being mentioned inside the user’s ChatGPT account via this prompt.

They also added a “Learn more on Glasp” link, hoping users would click back to the site after using ChatGPT.

What they found

After going to ChatGPT, the click-through to return to Glasp via their ‘Learn more’ button was weak.

Most users got what they needed inside ChatGPT and did not return to the site. So even when Glasp was mentioned or cited, the traffic was mostly lost.

The harder part is measurement. A prompt may affect what ChatGPT remembers inside one user’s own account, but it is much harder to know whether that changes what other users see. Any wider effect would be slow, unclear and difficult to prove. So the tactic is interesting, but still unproven.

How businesses can apply this tactic

If you did want to test out this tactic, a simple version is to add a “Summarise this in ChatGPT” button at the top of an article. When the user clicks it, the button opens ChatGPT with a short pre-filled prompt.

For example:

Summarise this article in ChatGPT

Then the prompt url includes the article text, or a short extract from it, so the user can see what is being sent.

This may help get your brand and article into more AI chats. But it should not replace the basics: clear summaries, Q&A sections, strong site structure, clean metadata and content that is worth citing.

What else may have influenced Glasp’s result

Glasp’s ChatGPT referral traffic grew from 517 daily sessions on 1 January 2026 to 19,129 daily sessions on 5 May 2026. The founders shared their traffic data, so the growth itself is well documented.

But when looking at the full picture, it is worth flagging the other factors that may have helped or influenced the result.

Here are some other elements that likely played a part.

They already had established domain authority

The Glasp founders are transparent that this experiment was not on a new website.

They reported that Glasp’s domain authority was around 69 to 70 at the time of the experiment. They had also spent years building content around highlights, knowledge management and YouTube Q&A, with backlinks from guest posts and other organic mentions. (2)

This meant AI tools were not being asked to understand or trust a brand-new site.

The takeaway is that AEO is likely to work better when the SEO basics are already in place.

ChatGPT’s own referral traffic growth may have added to the results

Another factor is that ChatGPT itself was growing as a referral source.

According to the Glasp founders’ later research paper, the pages that did not get the AEO changes still grew by 3.5x. That suggests some of the growth came from ChatGPT sending more traffic overall, not only from Glasp’s page changes.

But the pages that did get AEO changes still performed better. Glasp estimated that the AEO work added a further 1.8x to 2.3x lift. The paper describes this as “suggestive, not conclusive”, because there was not enough clean before-and-after data to prove it with certainty. (6)

They grew despite YouTube summariser tools getting more crowded

There is also a point in Glasp’s favour.

This was not an empty market. By 2025, YouTube and video summariser tools had become a crowded category. TechPP listed 10 YouTube video summariser tools in 2025, Recapio ranked 12 AI video summary generator tools, and Summarie AI said it had tested 15+ AI video summariser tools.

That means Glasp was growing while users had more options, not fewer.

It was competing with copycat tools, general AI assistants and dedicated video summary products. Yet ChatGPT referral traffic to its site still grew.

This makes the result even more interesting. Glasp was not growing in a static category. It grew while competition around YouTube summarisation tools was increasing rapidly.

Other AEO tactics working in the background

When I dug around further, I noticed that Glasp also had several other tactics working in its favour.

Some may have been intentional AEO work. Others look more like normal marketing and community-building. Either way, they are useful to flag because AI visibility is shaped by more than one page or one tactic.

1. They had active Reddit discussions

Glasp also had visible activity on Reddit, including its own subreddit. Users were reporting bugs, asking questions and voting on feature requests.

This may have helped because Reddit is one of the places AI tools can draw from when answering questions.

OpenAI announced a Reddit partnership in 2024, giving it access to Reddit’s Data API and bringing Reddit content into ChatGPT and other OpenAI products. So Reddit is not just another social platform. It is one of the public discussion spaces that can feed into AI answers.

For Glasp, this meant there were more natural mentions of the product outside its own website. Not polished marketing copy. Real user questions, issues, comparisons and requests.

That kind of content is useful because people often ask AI tools questions in the same way they ask Reddit: “What is the best tool for this?”, “Has anyone used this?”, “Does this work?”, “What are the alternatives?”

For businesses, posting in Reddit or niche community forums can help create the kind of natural, third-party mentions that AI tools may use as trust signals when deciding which source to cite.

2. They had brand mentions across the web

Glasp also had a strong presence beyond its own site.

The founders had appeared in guest posts, interviews, YouTube videos and other public content around AI, learning and knowledge management.

This may have helped because AI tools look for signals across the wider web, not just on your website. The more consistent mentions a brand has in relevant places, the easier it is for AI tools to understand what it does and where it fits.

For businesses, the takeaway is to build proof and mentions outside your own site: interviews, podcasts, guest posts, YouTube, LinkedIn, directories, review sites and partner pages.

3. Bing access may have helped with AI discovery

Bing is also worth checking as part of an AEO setup.

In one Glasp snapshot, Bing’s bot appeared in the crawl data. That does not prove Glasp had Bing Webmaster Tools set up, or that Bing caused the ChatGPT traffic growth.

But it is still relevant. Bing sits inside Microsoft’s search and AI ecosystem, including Copilot and Bing’s AI-generated answers. Microsoft and OpenAI also continue to have a close partnership.

Bing Webmaster Tools now also has an AI Performance report, which shows when your site is cited in AI-generated answers across Microsoft Copilot, Bing and some partner integrations.

For businesses, this makes Bing worth including in the AEO setup: submit the sitemap, check crawlability, review the AI Performance report if available, and use IndexNow so new or updated pages are easier for Bing to find.

Key takeaways

AEO is still early, and Glasp’s result should not be treated as a simple checklist that guarantees traffic.

But the founders’ experiments do point to a few practical things businesses can test.

1. Add a clear summary at the top of important pages

Add a short TLDR to articles, guides, landing pages and service pages.

Keep it short, but useful. Explain what the page covers, who it is for, and the main context an AI tool would need to understand it.

A good summary helps both readers and LLMs work out what the page is about quickly.

2. Add question-and-answer sections to key pages

Use the kinds of questions a real person would ask ChatGPT.

These can sit inside an article, at the end of a blog post, on service pages, or on landing pages.

The answers should be direct, useful and written in plain English. Repeat the main topic near the start of the answer so the context is clear.

3. Check server logs for missing pages AI tools are trying to fetch

Glasp found that ChatGPT was trying to access URLs that did not exist.

For businesses with structured content, this could be useful. Check your server logs for repeated 404s from AI bots. If the same missing page patterns keep appearing, they may point to content users are asking for and AI tools expect you to have.

4. Treat AEO as an authority problem, not only a page problem

Glasp had more than just AEO optimised pages. It already had domain authority, backlinks, a clear content base and mentions across the web.

That likely helped.

For most businesses, this means continuing the work that builds trust outside your own site: guest posts, interviews, reviews, directories, partner pages, YouTube, LinkedIn and useful community activity.

5. Build mentions where real users already talk

Reddit and niche forums may help because they create natural third-party mentions around a brand.

This does not mean spamming Reddit but you can contribute to Reddit discussions in the places where customers ask questions, compare options and describe their problems in their own words.

Those mentions may help AI tools understand what your brand does and where it fits.

6. Set up Bing Webmaster Tools

While most of us focus on Google’s Search Console, Bing matters too.

Set up Bing Webmaster Tools, submit your sitemap, check that Bing can crawl your site, review the AI Performance report if available, and use IndexNow for new or updated pages.

FAQs

What is Answer Engine Optimisation, or AEO?

Answer Engine Optimisation, or AEO, is the work of making a website easier for AI tools to understand, retrieve and cite in generated answers. In this article, AEO refers to tactics such as clearer summaries, question-and-answer sections, structured page content, server-log analysis and wider authority signals.

How did Glasp grow ChatGPT traffic by 37x?

Glasp’s founders reported that ChatGPT referral traffic grew from 517 daily sessions on 1 January 2026 to 19,129 daily sessions on 5 May 2026. They did this by studying server logs, improving page summaries, adding question-and-answer content, analysing missing URLs and making pages easier for AI tools to retrieve and cite.

Was the full 37x growth caused by AEO?

The 37x growth is real, but Glasp’s later research suggests the AEO work did not account for all of it. Some of the increase came from ChatGPT growing as a referral source overall. The founders estimated the AEO work added a further 1.8x to 2.3x lift, although they described the result as “suggestive, not conclusive”. (6)

What AEO tactics did Glasp test?

Glasp tested several AEO tactics, including clearer TLDR summaries, question-and-answer sections, Q&A structured data, server-log analysis, 404 content mapping and repeated brand mentions inside ChatGPT prompts. The founders also looked at how AI bots were crawling the site through Cloudflare logs.

Why do TLDR summaries help with AEO?

TLDR summaries help because they give AI tools a clear explanation of what the page covers near the top of the page. Glasp found that pages fetched most often by AI Assistant traffic had TLDRs of about 132 characters, while the least-fetched pages had TLDRs of about 14 characters.

Why do question-and-answer sections help with AEO?

Question-and-answer sections help because many people use AI tools by asking full questions. If a page already contains clear questions and direct answers, it may be easier for an AI tool to match that page to a user’s prompt and retrieve the relevant section.

Why are 404 logs useful for AEO?

404 logs can show which missing pages AI tools are trying to fetch. Glasp found that ChatGPT-User made 78,640 failed requests in 7 days, then looked for repeated URL patterns. When missing URLs appeared to point to pages that should exist, the team treated them as a demand signal and created new content.

Does Reddit help with AEO?

Reddit may help with AEO because it creates natural third-party mentions around a brand, product or category. OpenAI also announced a Reddit partnership in 2024, giving it access to Reddit’s Data API and bringing Reddit content into ChatGPT and other OpenAI products.

Should businesses set up Bing Webmaster Tools for AEO?

Bing is worth including in an AEO setup because it sits inside Microsoft’s search and AI ecosystem. Bing Webmaster Tools also has an AI Performance report, which shows when a site is cited in AI-generated answers across Microsoft Copilot, Bing and some partner integrations.

What should businesses test first for AEO?

The lowest-friction starting points are adding clear TLDR summaries, adding question-and-answer sections to key pages, checking server logs for repeated 404 patterns, and making sure search engines can crawl and understand the site. Businesses with existing authority, useful content and clean site structure are likely to be in a better position to benefit from AEO work.

---

References

1. KEO Marketing, “SEO Traffic Decline: Why 73% of B2B Websites Lose Visibility”, 2025.
2. Sean Ellis, Kei Watanabe and Kazuki Nakayashiki, “How Glasp grew ChatGPT traffic from 500 to 19,000 daily sessions in 4 months”, Growth with Sean Ellis, 22 May 2026.
3. TechCrunch, “Quora launches Poe, a way to talk to AI chatbots like ChatGPT”, 21 December 2022.
4. Ahrefs, “New Study: How Often Do AI Assistants Hallucinate Links? 16 Million URLs Studied”, 2025.
5. Glasp, “Hatching Growth #12: AEO in Practice: Memory Hacks, Prompt Injection Experiments, and Honest Results”, 20 October 2025.
6. Keisuke Watanabe and Kazuki Nakayashiki, “Disentangling Answer Engine Optimization from Platform Growth: A Log-Based Natural Experiment on ChatGPT Referral Traffic”, arXiv, 4 June 2026.
7. TechPP, “10 Best YouTube Video Summarizers in 2025”, 2025.
8. Recapio, “12 Best AI Video Summary Generator Tools for 2025”, 2025.
9. Summarie AI, “AI Video Summarizer Comparison 2025”, 2025.
10. OpenAI, “OpenAI and Reddit Partnership”, 2024.
11. Bing Webmaster Tools, “AI Performance in Bing Webmaster Tools”.

---

Sign up to the Agentic Formula newsletter for the next research-backed AI case study, straight to your inbox.

---