Quick Answer
Glasp grew daily ChatGPT traffic from 517 sessions to 19,129 in four months, which is where the 37x comes from. Against a control group, the untreated pages on the same domain that grew 3.5x with none of the work, Glasp's own analysis puts the effect of its changes at about 1.8x to 2.3x. The honest reading is roughly 2x, and Glasp calls even that suggestive rather than proven.
Most AEO case studies stop at the biggest number they can screenshot. Glasp, the web highlighter behind the YouTube Summary with ChatGPT extension, published a big one and then spent half the write-up taking it apart.
That second half is the rare part, and it's worth copying before any tactic. If you run a small site, the multiple you report decides what gets funded next quarter.
This covers how Glasp's 37x becomes about 2x in five steps, the five tactics that transfer to a small site, and the one precondition Glasp had that you probably don't.
How Glasp's 37× Shrinks to About 2× Against a Control Group
The 37x shrinks to about 2x because each step of Glasp's own accounting strips out growth the work did not cause, and the last step is a control group. Kei Watanabe and Kazuki Nakayashiki, Glasp's co-founders, laid the cascade out in Hatching Growth #17 and published the full model as a paper on arXiv.
| Step | Multiple | What it compares |
|---|---|---|
| Two single days | 37x | One day's sessions against another's |
| Monthly totals | 5.7x | All ChatGPT referrals, January against May |
| Untreated pages | 3.5x | The control group, grown by ChatGPT alone |
| Treated pages | 6.1x | The YouTube pages that got the work |
| Time-series model | 1.82x | The jump in the treated-to-control ratio |
Read top to bottom, the table is the distance between a screenshot and a claim. Only the last row is something Glasp's work can own, and the steps below show how each row gets there.
1. Two Single Days Produced the 37× Figure
The headline compares 517 daily sessions from chatgpt.com with 19,129 about four months later. Those are human click-throughs recorded in GA4, not bot hits, and Glasp notes that ChatGPT's mobile apps strip referrers, so the figure probably undercounts.
The weakness is the endpoints, not the tracking. A single day is volatile, and Glasp reports that the sustained late-May level was under half that peak.
On your own numbers, never compare two days. Compare two equal windows of at least a week, and decide in advance which windows you'll use, the way Glasp pre-committed to a 10,000-session threshold before it pitched the story.
2. Monthly Totals Shrink the Multiple to 5.7× Growth
Aggregated by month, total ChatGPT referrals grew 5.7x from January to May. That's still a strong result, and it's already a long way from 37.
Monthly totals smooth out the single-day spike, but they still mix your work with everything else that happened in those months. The common mistake is to stop here and report 5.7x as the effect of the work, with no control group to separate the two.
3. Untreated Pages Grew 3.5× With No Work at All
The control group is what exposes the tide. Glasp applied its changes only to its corpus of YouTube question-and-answer pages, and the rest of glasp.co never got them.
Those untreated pages still grew 3.5x over the same window, while treated pages grew 6.1x. The 3.5x is ChatGPT's own growth, and the gap between the two lines is the only part Glasp's work can claim.
Site-wide work such as performance tuning touched both groups, so any help it gave the control group makes the true effect larger, not smaller. The founders put the point plainly: "If someone quotes you a raw AEO multiple on a channel that grew 3.5x untouched, they are taking credit for the weather."
4. An Interrupted Time Series Put the Effect Near 1.8×
An interrupted time series asks whether a trend jumps at the moment an intervention starts. Glasp ran one on the weekly ratio of treated to control traffic and found a discrete 1.82x jump aligned with the January rollout, with a 95 percent confidence interval of 1.31 to 2.54.
Filtered to engaged sessions only, the same model read 2.27x. That's where the "about 1.8x to 2.3x" range comes from, and it's the number worth repeating.
You can run a rough version with a spreadsheet. Plot the weekly ratio of the pages you changed to the pages you didn't, mark the week you shipped, and look for a step rather than a slope.
5. A Placebo Test Kept the Result Suggestive
Glasp still would not call the 2x proven. A placebo permutation test across the weeks before the rollout returned p=0.16, meaning jumps nearly that large sometimes appeared by chance in a short, noisy pre-period.
So the paper says the effect is suggestive, not conclusive. That sentence is the part most case studies leave out, and it's the reason this one is worth reading.
What Glasp Changed Before the Control Group Measured It
Glasp changed how it chose, named and summarized pages, using its own server logs rather than a vendor's estimate of what ChatGPT cites. Comparing the changed pages with the control group, the rest of glasp.co, then measured what those changes were worth.
The logs came from Cloudflare AI Crawl Control, which Cloudflare lists as available on all plans. Glasp sorted its rewrite queue by the AI Assistant category, which counts real-time fetches such as ChatGPT-User retrieving a page mid-answer.
- It compared its most-requested pages with its least-requested ones and found a summary gap: about 132 characters against 14.
- It merged two URL patterns that served identical content, because bot traffic was splitting across the duplicates.
- It rewrote titles into the question a user would type, and thin summaries into 2-3 sentence answers, starting with about 12,000 pages ranked by live demand.
- It built an SEO Guard: pages with 100 or more Google clicks in 180 days were locked, and pages with no impressions and no AI requests were unpublished.
One measurement stands out. Weekly AI bot requests stayed roughly flat at 2 to 3 million while human sessions climbed, so the work changed how often pages got chosen, not how often they got crawled.
Five Glasp Tactics a Small Site Can Run With Free Tools
Five parts of Glasp's method transfer to a small site, because they rely on data you already own rather than on scale. None of them needs a control group to start, though you'll want one before you claim a result.
1. Snapshot AI Crawl Control Before the Data Ages Out
Cloudflare's dashboard only looks back a short way: Glasp reports 24 hours on the free tier and seven days on its own plan. Glasp pulled the data by API every week and stored it, which turns a rolling window into a history.
A weekly cron job writing to a spreadsheet does the same job on a small site. After a month you can see which pages ChatGPT-User fetches and whether that changes after an edit.
2. Treat AI Bot 404 Errors as a List of Missing Pages
A 404 from ChatGPT-User means the model guessed a URL on your domain, tried to fetch it mid-answer, and found nothing. Glasp cites an Ahrefs study that found AI assistants send visitors to 404 pages 2.87 times as often as Google Search.
In one seven-day window Glasp logged 78,640 unsuccessful ChatGPT-User requests, though that bucket also held 403s, 429s and server errors. Filter to the 404s, group them by URL pattern, and the repeats are a list of pages people already asked for.
3. Audit Your Firewall for Rejected ChatGPT-User Requests
At the worst point, Glasp found its own security rules rejecting about a quarter of ChatGPT-User requests, and a rate limit refusing ChatGPT's retrieval intermittently. Nothing looks wrong in a browser when this happens.
Check your WAF and bot rules for 403 responses to AI user agents before you touch a single page. It's the cheapest fix on this list. Rendering is the other quiet failure, since a React site can hand crawlers that read raw HTML a near-empty page, as covered in why AI crawlers read raw HTML.
4. Read Grounding Queries in Bing Webmaster Tools
Bing's AI Performance report shows citations in AI-generated answers and the grounding queries, the phrases the AI used when it retrieved your content. Glasp calls it the AEO dashboard nobody reads.
If the 404 list tells you what should exist, grounding queries tell you what already works. Both are free, and Google Search Console fills in the queries earning impressions on Google.
5. Create New Pages Only From Observed Demand
Glasp ranks this above every rewrite it made. Pages Glasp created in bulk from March to May earned any Google impressions 0.3 to 0.8 percent of the time, and pages created from observed demand in June hit 13.4 percent.
The generation prompt had not changed, only the choice of which pages deserved to exist. Starting from the question a buyer actually asked is also where Better Marketing's AI SEO service begins, because an assistant quotes an answer and skips a keyword page.
For content marketing for tech companies, the lesson is uncomfortable. A well-written page nobody asked for still earns nothing, so the editorial calendar should come from your logs and your Search Console, not from a keyword list.
Where Glasp's Playbook Breaks on a Low-Authority Domain
Glasp's playbook breaks on a low-authority domain because it assumes discovery is already solved. Glasp says its domain authority was 69 to 70 when the experiment began, and that it does not know what happens from a cold start.
| Factor | Glasp | A 40-page startup site |
|---|---|---|
| Domain authority | 69 to 71 | Near zero |
| Pages to work with | 400,000+ | Dozens |
| Bottleneck | Being chosen | Being found |
| Control group | Untreated rest of the site | Rarely large enough |
Tombstoning half a corpus, which Glasp did to more than 200,000 pages, is no strategy when you have forty pages worth keeping. A 12,000-page rewrite queue doesn't translate either. For content marketing for tech companies at this stage, a handful of pages built from real questions beats a corpus built from a keyword export.
A small site also rarely has enough traffic to split into treated and untreated groups. The fallback is a dated change log and a before-and-after on equal windows, reported as a hint rather than a control group result. For a zero-authority starting plan, the 30-day AEO plan for startups covers the order of work.
How to Read Any AEO Case Study With a Control Group in Mind
Read any AEO case study by asking what the multiple compares and whether a control group absorbed the platform's own growth. Glasp's write-up answers every question below, including the ones its own headline fails, which is why it makes a useful benchmark.
- Check the endpoints: two single days, or two equal windows chosen in advance.
- Check the unit: bot requests, human sessions, or citations, since Glasp's bots stayed flat while sessions rose.
- Look for an untreated group on the same domain over the same weeks.
- Look for an interval or a placebo test, not only a point estimate.
- Ask what happened to the channel that was not the target; Google clicks to Glasp's rewritten pages fell about 25 percent, in line with the site-wide trend, with no deindexation.
Context matters too, because AI referrals are still a small share of most traffic, as the Cloudflare Radar referral numbers show. Any AI SEO service that reports a raw multiple without a control group is reporting ChatGPT's growth and calling it theirs.
The same test applies when you hire a B2B demand generation agency. Ask what share of the lift the untouched part of your pipeline also saw, and treat the difference as the result.
Measure Against a Control Group Before You Claim a Lift
Glasp's 37x is real, and so is its roughly 2x. The work lives in the gap a control group reveals, not in the screenshot.
Copy the free parts first: snapshot your AI crawl logs, fix what your firewall rejects, and build pages from 404s and grounding queries. Then report the change against an untouched set of pages.
Glasp left one question open, what AEO traffic converts into, and it is the question a B2B demand generation agency should answer with numbers. Better Marketing works on that part, turning the trust a citation earns into first customers.