We already wrote the theory: structured content, backlinks, third-party mentions, consistency. Good advice, but advice is easy to write and hard to verify. So instead of publishing another checklist, I ran the playbook on a real product I own, roastme.gg, and logged the actual before-and-after with real API transcripts instead of screenshots I could have cherry-picked.
roastme.gg is a small, silly, paid product: you submit your X handle or a website, pick a dollar amount from $1 to $1,000+, and an AI writes a public roast of it that gets posted to a leaderboard ranked by how much you paid. It launched on Product Hunt and I posted it on Show HN the same week this test ran. Low stakes, real traffic, real Stripe payments, exactly the kind of indie product this blog is for. (Here's the honest revenue and pricing breakdown if you want the numbers instead of the citation test.)
Here's what I actually did, what the AI search engines actually said before and after, and the one thing I got embarrassingly wrong along the way.
The setup: querying GPT-5 with real web search, not guesswork
Screenshotting a single ChatGPT answer and calling it a case study is how most of these posts go wrong. A single response is one sample from a system that reformulates its sources every time. So I used the OpenAI Responses API with the web_search tool, the same grounded search mechanism the ChatGPT product uses, and ran the exact same five prompts before shipping any changes and again after.
The five prompts, written the way an actual buyer would type them, not SEO-keyword-stuffed:
- "What's a website where you can pay to get publicly roasted by AI?"
- "AI roast my twitter account, is there a site for that?"
- "Best AI roast generator websites 2026"
- "site where people pay money to get insulted by AI and posted publicly"
- "what is roastme.gg"
Query 5 is the control. If a model can't describe a product correctly when asked by name, nothing else matters. Queries 1 through 4 are the real test: does the product show up when nobody already knows it exists?
Round one: the homepage had almost nothing to cite
Before touching anything, I read roastme.gg's own homepage the way a crawler would. The entire visible body text was two lines: "Get roasted. Publicly." and "Ranked by how much it hurt your wallet." Everything else, the five pricing tiers, how the wall ranking worked, the self-roasts-only policy, existed only inside an interactive submission widget or a meta description tag. Structured data was a single generic WebSite schema. No FAQ, no pricing markup, no answer-first content anywhere.
That's not a rare mistake. It's the default state of most indie landing pages: everything the product does lives in a hero line and a screenshot, not in sentences a machine can extract and repeat.
Running the five prompts against that homepage:
| Query | Cited roastme.gg? | What showed up instead |
|---|---|---|
| Pay to get publicly roasted by AI | No | KindRoast, RoastMyPic.ai |
| AI roast my twitter account | No | profiled.lol, Roast That Face, The Roast Bot, Voicemoat, Contagent |
| Best AI roast generator websites 2026 | No | roastai.app, Mugzilla, TipsyText, roast-me.net |
| Pay money to get insulted by AI, posted publicly | No | Your Music Sucks, Roast My Wallet, ShameMySite, Silicon Social |
| what is roastme.gg | Yes | Correctly described, cited the homepage |
One out of five, and the one hit only worked because the query already contained the brand name. For every generic query, a different competitor showed up, and three of them (Roast My Wallet, ShameMySite, Silicon Social) turned out to be near-identical products: pay-to-rank public walls. That's the actual competitive set for citation share in this niche, not "AI roast tools" broadly.
What I shipped, in order of how mechanical it was
1. A static "how it works" section and an FAQ page. Instead of bolting FAQ content onto the homepage (which I initially did, then pulled back off because it made the homepage feel cluttered), I moved it to a dedicated /faq page: plain-text pricing tiers with real dollar ranges, how wall ranking works, the self-roasts-only policy, and six Q&A pairs including one that surfaces a real live number pulled from the database ("The Wall currently has 4 public roasts, with the top spot going for $20").
2. FAQPage and Service JSON-LD. The FAQ page got FAQPage schema wrapping each question. The root layout got a Service schema with an Offer entry per pricing tier, minPrice/maxPrice and all, so the pricing structure is machine-readable, not just prose. Worth being honest here: Google retired FAQ rich results in May 2026 and its own May 2026 guidance says structured data isn't required for generative AI search. I added it anyway because it's nearly free and it still helps Bing and Perplexity parse the Q&A, but I did not expect it to be the thing that moved the needle.
3. Structured data on every individual roast page. Roast pages are the highest-volume, most-shared URLs on the site, and they had zero JSON-LD, just basic Open Graph tags. Added CreativeWork (the roast text, tier, publish date, and for website targets, an about entity pointing at the roasted site) plus a BreadcrumbList. Small, mechanical, but these are the pages that actually get shared on X, so they're the ones most likely to get crawled.
4. Existing OG/meta gaps. Missing og:locale, twitter:site, theme-color, and a web manifest. Not an AI-visibility fix specifically, just overdue hygiene that happened to get bundled into the same week.
None of this touched off-site presence. That was deliberate: I wanted to isolate what on-page structure alone does before adding the harder, slower variable of third-party mentions.
What broke along the way (the embarrassing part)
While pulling pricing data into the new FAQ page, I went looking at how individual roast URLs get their slugs and found a bug that had nothing to do with AI visibility and everything to do with basic correctness: the slugify() function stripped a leading @ for X handles but never stripped https:// for website targets. Submit https://example.com and the public URL became roastme.gg/roast/https-example-com instead of roastme.gg/roast/example-com.
Only one live row had it, caught with a read-only SELECT before touching anything. I fixed the function, updated that one row's slug with a logged rollback query ready in case anything looked wrong, and added a permanent redirect from the old broken URL so it wouldn't 404 if it had already been shared. Small bug, but it's a good reminder that a GEO push is a fine time to actually read your own generated URLs, because nobody else will flag them for you.
Round two: the same five prompts, right after deploy
I re-ran the identical five prompts within minutes of the deploy going live. If you're expecting a dramatic before/after screenshot, I don't have one, and I'd be skeptical of anyone who claims they do on this timeline. The generic queries (1 through 4) still didn't mention roastme.gg. Same competitors surfaced, in a slightly different mix each run, which is its own finding: this niche is fragmented enough that even the incumbents rotate depending on how the model formulates its search.
But query 5, "what is roastme.gg," changed in a way worth noting. Before, it cited the homepage and described the product in general terms. After, it cited /faq specifically and quoted the pricing tiers almost verbatim: "Pricing tiers run from Mild ($1-$4) up to Scorched Earth ($1,000+); higher spend = harsher roast and higher rank." That's not a new mention, it's a better one, sourced from a page that didn't exist an hour earlier, pulling structured facts instead of paraphrasing marketing copy.
That's the honest, unglamorous finding: on-page structure gets crawled and used fast when a model already knows to look for you by name. It does nothing measurable for cold discovery on the same timeline, because cold discovery runs on a completely different signal.
What I'd do differently starting today
If I were doing this from scratch instead of retrofitting it onto a product that already had traffic, I'd change the order of operations:
- Ship the FAQ page and pricing structure before launch, not after. It took an afternoon. There's no reason it should have shipped weeks after Product Hunt instead of the same day. Indexing lag is real, so every day the structured content isn't live is a day it isn't in the queue.
- Treat the homepage as a landing page for humans and put the answer-first content somewhere dedicated for machines and search. I originally tried cramming the how-it-works and FAQ sections directly onto the homepage, then pulled them into
/faqbecause the homepage started reading cluttered. That split turned out to be the right call twice over: a cleaner homepage for the person paying to get roasted, and a URL I can point directly at when a query is clearly informational. - Log the baseline before making any changes, not after. I got lucky here because I happened to run the five prompts before shipping anything. If I hadn't, I'd have no way to separate "the model always knew this" from "the model learned this today," and every case study without a real baseline is just a screenshot with vibes attached.
- Don't wait for citation data to start the backlink work. The Show HN post went out the same week as the on-page changes, not after waiting to see if the on-page work alone would be enough. Given the 85% third-party stat above, waiting on generic-query citations before starting outreach would have wasted the slowest part of the timeline doing nothing.
Why the generic queries didn't move (and won't, soon)
Three things from the wider research back up what I saw, and they're worth knowing before you run this same test on your own product and get discouraged by day-one results:
Indexing has a real lag. ChatGPT's search backend runs on Bing's index and typically takes one to three weeks to pick up new content. Perplexity is faster, often days. Google AI Overviews follows normal Google indexing, four to eight weeks for a meaningfully changed page. A same-day re-test on generic queries is checking a clock that hasn't ticked yet.
Third-party mentions are the actual lever, not owned content. Multiple 2026 analyses put the split at roughly 85% of AI citations coming from third-party pages, publishers, forums, review sites, directories, versus your own domain. One study found brands are 6.5x more likely to get cited through a third-party source than their own site. Reddit alone is the second most-cited domain in Google AI Overviews behind YouTube. This is the same reason dofollow backlinks matter for indies beyond classic SEO: a listing that gets you a real, indexed, third-party mention is worth more to AI visibility than another paragraph on your own homepage.
llms.txt is not a citation lever, full stop. roastme.gg already has one; I left it in place. But independent studies covering 300,000+ domains in 2026 found no measurable citation effect, and Google's own documentation update in June 2026 confirms Search and AI Overviews simply ignore it. If you're choosing where to spend an afternoon, spend it on the FAQ page and the third-party listing, not the llms.txt file.
Specific, structured claims outperform generic marketing copy when they do get read. Research out of Princeton and Georgia Tech on generative-engine optimization found that adding inline citations to primary sources improved citation rates by 40%, adding specific statistics improved rates by 37%, and named expert quotes added 22%. That tracks with what I saw on query 5: the model didn't just mention roastme.gg more confidently, it quoted a real number ("$1-$4" instead of "cheap tier") because a real number was finally there to quote.
Frequently asked questions
Does llms.txt actually help you get cited by AI?
No, not on the current evidence. It's a nearly-free file to add and it isn't harmful, but three separate 2026 studies covering hundreds of thousands of domains found zero measurable effect on citations, and no major AI company has confirmed reading it in production. Spend your time on visible, structured page content instead.
How long before AI search visibility changes actually show results?
Perplexity can pick up new content within days. ChatGPT search typically takes one to three weeks because it runs on Bing's index. Google AI Overviews can take four to eight weeks. A test run the same day you ship is really only useful for confirming the new page got crawled at all, not for judging citation share.
Do backlinks still matter if AI doesn't rank pages the same way Google does?
Yes, arguably more, not less. AI systems lean on third-party mentions as a trust signal precisely because they can't verify your own marketing copy. A dofollow link from an indexed, relevant page (a curated directory, a real guest post, a Show HN thread that gets discussed) does double duty: it's still an SEO signal, and it's the kind of external mention that gets your product described accurately somewhere you don't control.
What's the single highest-leverage thing to do first?
Read your own homepage as plain text with the styling stripped away. If you can't answer "what is this, what does it cost, how does it work" from the words alone, no amount of schema markup or llms.txt will fix that. Fix the words first. Structured data is a multiplier on content that already exists, not a substitute for it.
Should structured data live on the homepage or a dedicated page?
Either can work, but a dedicated FAQ or pricing page gives you a URL you can point to directly, keeps the homepage focused on the person deciding whether to buy, and makes it obvious later which page is actually getting cited. Roast pages on roastme.gg also carry their own CreativeWork and BreadcrumbList schema rather than leaning on the homepage alone, since those are the pages that actually get shared.
The honest summary
On-page fixes, a real FAQ page, structured pricing data, JSON-LD on the pages that actually get shared, changed how confidently and accurately AI search engines describe roastme.gg when someone already asks about it by name. They did nothing measurable for cold discovery in the same 24 hours, and based on the indexing timelines above, they're not going to for a few more weeks. That's not a failure of the tactics. It's just an honest timeline nobody selling a "get cited by ChatGPT" service is incentivized to tell you.
The next real test is the slower one: whether the Show HN post, the directory listings, and whatever third-party mentions accumulate over the next month actually move the generic queries. I'll update this post with round three once there's enough time on the clock for it to mean something.
If you want the same starting checklist we used before running this test, read the AI visibility guide first. And if you're listing your own product to start earning the third-party mentions that actually move AI citations, list it on makers.page for free, roastme.gg is already up there too.
Related reading: