Divij Rakhra
On September 22, a website that was 33 days old and had 3 backlinks started showing up inside ChatGPT answers.
Ten times on day one. Fourteen the next day. By September 29 it had been pulled into ChatGPT answers 92 times and cited 113 times.
Over those same eight days, across the same tracked prompts, clutch.co got 37 retrievals and 8 citations. Clutch has 48,372 referring domains. This site has 3.
It's an exact-match domain (EMD): a domain that is literally the search query, with a ranked directory on it. It's the second one I've built. The first one did the same thing back in July, and I wasn't sure if that was a fluke. This post is both experiments, every number, and the exact playbook, so you can run it yourself.
The short version: all good SEO is good AEO, but not all good AEO is good SEO. Google still wants authority before it shows you to anyone. ChatGPT's retrieval layer mostly wants the page that looks exactly like the answer. That gap is the whole opportunity.
Experiment 01A UX agency directory
On July 2 I bought an exact-match domain for the head term in the UX agency category. I built a ranked directory on it with Claude Code in an afternoon: a scoring rubric, a top list, a review for every agency, and a source link on every claim. I shipped it and did zero outreach.
Google did what Google does with a brand-new domain. It indexed it in three days and then kept it on page six and beyond. In 67 days: 18,776 impressions, 16 clicks, average position 68.6.
ChatGPT did something completely different.
Out of the 1,000 domains that ChatGPT, Perplexity and Google AI Overviews retrieved when answering questions in the category, a 2-month-old domain with six backlinks ranked #73. Clutch, with 48,372 referring domains, was in the same table.
That was one site. One site can be luck. So I built a second one, changed a few things based on what I'd learned, and measured it properly.
Experiment 02A narrower directory
This time I went narrower: one buyer segment instead of the whole category. The site launched on August 20. Everything that shipped, with dates:
The lag is the part I'd tell anyone copying this to hold their nerve through. The first site took about six weeks from its first retrieval to getting our agency named in answers. This one took 33 days to get retrieved at all, and then it went vertical. Two weeks after the directory went live, it flipped on.
The resultWhat it did for our agency
Getting a directory cited is nice. Getting our agency recommended is the point. So I pulled every ChatGPT answer for the 11 prompts where EMD #2 was being retrieved: 7 days before it started getting cited, and 8 days after. 165 answers in total.
When the EMD was one of the sources, ChatGPT named our agency in 38 of 52 answers. The EMD was the very first source cited in 13 of them.
“ux agencies for ai companies” is the one I'm proudest of. We built one page on EMD #2 for exactly that question. That page got retrieved 8 times and cited 29 times, 3.6 citations per retrieval, the highest of any page on either site. We were named in all 8 answers that week, and listed first in 6 of them.
The pagesWhich pages did the work
On the first site, 66 of 67 retrievals went to the homepage. The ranked list was the entire asset. On the second site I built pages shaped like the questions buyers actually type, and the citations spread out.
The mechanismWhy this works
When you ask ChatGPT for an agency recommendation, it doesn't just answer from memory. It rewrites your question into a few search queries, retrieves pages, and then writes the answer from them. Peec logs those rewritten queries. These are the most common ones from answers that cited EMD #2:
Look at the words. The rewritten query is almost literally the domain name. When the query is “B2B SaaS product design agency” and there's a site whose domain is those exact words, with a page title that repeats them, that's a near-perfect match on the host and the title.
Google weighs that match against links and history, and a new domain loses. The retrieval step inside ChatGPT is much more willing to take the match at face value. Then a second thing kicks in: Peec classifies both sites as Editorial, not Corporate. A ranked comparison reads to a model as a reference, and references get quoted.
The playbookIf you want to copy it
This is what I'd do again, in order. None of it needs a link budget.
- Buy the query, not a brand. The domain should read like the question your buyer asks. Ours matches the rewritten query almost word for word.
- Lead with the answer. The first paragraph names #1, #2 and #3 and says why, in plain sentences. That's the part that gets lifted.
- Publish the rubric. Weighted criteria, a score per agency, a total, and a source URL on every factual claim. A model trusts a list it can check.
- Keep it short and dated. A top 7, a month in the title, a stated review cadence. Then actually re-score when you say you will.
- Emit it as structured data. Rating, review and per-criterion scores on every entry. An llms.txt data dump, plus llms-full.txt with the whole corpus.
- One page per buyer prompt. Track prompts in a tool like Peec, then build a page shaped exactly like each one. Our AI page became our most-cited URL.
- Add depth once the list works. A full directory with a page per agency made the site look like a reference. Retrieval started two weeks later.
- Measure citations, not rankings. Your rank tracker will say this failed. Watch ChatGPT, where both sites broke through, and give it 4 to 6 weeks.
NextWhat's next
Two sites, two niches, same result. That's enough for me to stop calling it a fluke.
Next I'm building head-to-head comparison pages and running the same before-and-after read on them. I'll post the numbers here.
AI search data. Peec AI. Tracked prompts run daily across ChatGPT, Perplexity and Google AI Overviews. A retrieval is one answer that pulled at least one URL from a domain. A citation is an inline citation in the answer. Citation rate is citations divided by retrievals. EMD #1 figures are from the Sep 9, 2026 read (Jun 1 to Sep 8, 2,829 answers, 71 prompts). EMD #2 figures are from the Sep 30, 2026 read. The before/after comparison uses all 165 ChatGPT answers for the 11 prompts that retrieved EMD #2, Sep 15 to 29. Other things changed in that window too, so treat the lift as a strong correlation, not a lab result.
Search and authority data. Google Search Console API. DataForSEO Backlinks API, pulled Sep 9, 2026. DataForSEO domain rank is a 0 to 1000 scale, not Ahrefs DR. Domain names and author names in screenshots are redacted.