AI answers · Perplexity and Gemini

An AI answers playbook for Perplexity and Gemini

Short answer: on 14 August 2026 Reddit's share of ChatGPT Search citations fell from 3.83% to 0.52%. The write-up is ChatGPT stopped citing Reddit. ChatGPT naming you from a Reddit thread sits outside this playbook. The weekly routine here is for Perplexity, Gemini and other AI answers: the buyer's question, a thread that already exists, a comment specific enough to lift. The GEO hub and the ChatGPT page are the map. The long ChatGPT guide is what changed on your own site. Karmy finds the thread and drafts the comment. You post it. A karmito is coming soon. Nobody can force a model to say your name.

Citation is a side effect, and that is the whole point

An AI citation, in the only sense that pays, is this: a stranger asks an assistant "what should I use for X," and your product is one of the names in the answer, with a reason that sounds like a person said it. Not a badge. Not a screenshot in a pitch deck. A sentence in someone else's buying process.

This page is the commenting habit we run when you recheck Perplexity, Gemini and other AI answers. On 14 August 2026 ChatGPT's Reddit share fell from an average 3.83% of ChatGPT Search citations (18 July to 7 August) to 0.52% (14 to 17 August). That post has the numbers, and the ChatGPT guide is the page about your own site: lead with the answer, add FAQ and article markup, do not invent ratings. The GEO hub and the ChatGPT worked example sit beside this routine. This page will not repeat the schema lecture, and it will not tell you a Reddit comment puts your name in ChatGPT.

Here is the opinion we are willing to put under our name. Chasing citations as a growth hack produces the comments models should ignore: a fresh account, a link, a superlative, the same paragraph in six subs before lunch. Those comments get removed. A removed comment is not a source. The playbook is mostly a list of ways to stay boringly useful long enough that a real thread still contains your name when an assistant looks.

Write the question the way a buyer types it

Before you touch Reddit, write three prompts in a note. Not keywords. Prompts.

  1. The job: "best way to log espresso shots at home" or whatever your actual job-to-be-done is.
  2. The switch: "alternative to [the tool people already name]."
  3. The complaint: "[that tool] is too expensive / too cluttered / missing the one thing."

Those three lines are your citation targets. If an assistant's answer to them never overlaps with a Reddit thread you can contribute to, you do not have a citation problem. You have a product-language problem, and no bot fixes that. Say the thing the way the buyer says it. The Brewlog example we use on how it works is "dial in espresso, log recipes," not "AI-powered beverage intelligence."

Ask the three prompts in Perplexity and Gemini this week. Write down every product they name. That list is your competitive set, whether you like it or not. ChatGPT is the engine that dropped Reddit, so a missing name there is not the scoreboard for this playbook. You are not trying to "beat SEO." You are trying to be a plausible member of the list those two answers already use, for a reason a skeptical reader would accept.

Only four kinds of thread are worth a draft

Karmy scores for intent. In public language, on the SaaS-on-Reddit guide, that means a need, a tool request, or a competitor complaint. We would add a fourth and then stop. If a thread is not one of these, skip it, even if it is huge.

Skip launch megathreads, meme posts, and any subreddit whose rules say no products, no links, no founders. A brilliant comment in a sub that removes it is a diary entry. Also skip threads where the useful answers are already a day old and buried. The first-hour rule is not a growth slogan. Early comments get read. Late comments get archived. Assistants are not doing you a favor by quoting comment number 400.

What Karmy is for, precisely. You add the product, the site and a few use cases. We watch for fresh threads in that shape and send them to Telegram with a draft, plus a sense of when the thread is still early enough to matter. We read the subreddit rules before the draft. We do not send you "everything mentioning your keyword." If you want that firehose, a keyword alert is the honest tool, and we compare a few of them below.

The comment has to be liftable

Models quote structure and specifics. They do not quote vibes. A comment that can survive being lifted into an answer has five properties, and we will be picky about them because vague "add value" advice is how people write ads with extra adjectives.

  1. The answer is in the first two lines. If someone only reads that far, they got the recommendation or the method.
  2. One detail you could be wrong about. A ratio, a step, a tradeoff, a thing you tried that failed. Vague praise is uncitable because it could describe any product.
  3. The other options are allowed to be good. "Notebook still wins if you want paper. An app wins if you want the last three shots next to the bag." A comment that only has one name reads as an ad, to humans and to anything summarizing humans.
  4. Your product is a clause, and you say you built it. Same sentence. Not a footer. Not a fake "a friend uses." Disclosure is not a legal costume here. It is why the comment is allowed to stay.
  5. No bare link in line one, and no link at all if the rules flinch. The thread can mention you without a URL. Your site can be found. A link is what moderators and spam filters look for first.

Here is the shape, using Brewlog, which is our example product and not a case study. Imagine the thread: "What's everyone using to dial in espresso? Notes app isn't working."

I log dose, yield, time, and the one thing I changed. The numbers alone never taught me anything. Writing the change next to the taste did. A notebook is honestly better if you pull two shots a week. I got annoyed enough to build a small logger (Brewlog, I'm the one who made it) because I wanted the last shots sitting on the coffee bag, not in a graveyard note. If you already like a spreadsheet, don't switch.

That comment can be shortened by an assistant into "people who log the change next to the taste mention Brewlog, and also say a notebook is fine." That is a citation. "Check out Brewlog, the best espresso app of 2026" cannot be shortened into anything except an ad, so it gets dropped, downvoted, or removed. We would rather you post the first one twice a week than the second one daily.

Edit every draft, including ours. The sentence only you can write is the one that keeps it from sounding generated. If you cannot add that sentence, do not post. A correct-but-generic reply trains both the subreddit and the model to skip you.

A dead account cannot be cited

This is the unglamorous half, and it is where playbooks usually lie. If the account is new, you are not in the citation business yet. You are in the not-getting-removed business. Our warm-up guide is the schedule for an account you run yourself: the first week is comments with no links, links come later, promotion stays rare, and the account should be a month old before it behaves like a marketer. The shadowban page is how you notice Reddit is hiding you. Check logged out. If the comments are gone for everyone else, stop posting and fix that before you write another "citable" paragraph into the void.

A karmito is coming soon: a Reddit account Karmy runs for you. The Telegram bot, the webhook and the MCP still do not post to Reddit. Do not blend a karmito with the 30-day guide and call it one plan. On Starter, Growth and Scale the draft does not click submit. On Scale, a weekly check of ChatGPT, Perplexity and Gemini. If someone sells you a promise that ChatGPT will name you from Reddit, they are not describing our pricing page or the August drop.

Two rules we will repeat until they are boring. Do not paste the same comment across subreddits. Do not put the link in a sub that forbids it, "just this once." The comment that gets removed cannot be quoted. The comment that stays up as a pure answer, with your name nowhere in it, still builds the account that will be allowed to speak next month. That is not a consolation prize. That is the work.

One original post, and only when you have the scar

Answering is the default. An original post is justified when you have a result, a failure, or a comparison you actually ran. "I tried three ways of logging shots, here is what I kept" can live for months and match the prompt you wrote at the top. "Launching today, please support" matches nothing an assistant is asked.

On Growth, Karmy can hand you ready-to-post ideas in that spirit. They are still ideas. You owe the post a real detail, the same way you owe a comment one. Read self-promotion without being that guy before you publish. If the post is you talking about you for twelve paragraphs, do not publish it, no matter how polished the draft looked in Telegram.

Reply under your own post. A thread the author abandons looks like a billboard. A thread the author answers looks like a conversation, which is the only Reddit artifact worth citing.

The Tuesday routine

This is the whole playbook once the account is allowed to speak. It should fit in the cracks of a build day, not become the job.

  1. Open Telegram. Look at what Karmy sent. If you are still doing this by hand, search only your three prompts, in the subs where your buyers already talk, and ignore everything else.
  2. Take two threads, not ten. If it is not one of the four shapes above, archive it in your head and move on. Scarcity is the feature.
  3. Edit the draft. Add the detail only you know. Cut the sentence that sounds like a tagline. Confirm the rules one more time if the sub is strict.
  4. Post from your account. Then stick around long enough to answer the first reply. The downvote guide is the tone check if you are unsure why a comment died.
  5. Once a week, rerun the three prompts in Perplexity and Gemini. Keep a note: date, who got named, whether you were one of them. Believe the answer box first. The check is not a lab report with a prompt quota, and it is not a promise about ChatGPT.
  6. Once the account is old enough, publish a single original writeup if you have something true to say. Not on a schedule you cannot feed.

Ten days free, no card to start. Then Starter is €9 a month for one mission, Growth is €19 a month for up to five missions, and Scale is €39 a month for up to 20 missions. On Growth and Scale the same opportunity can leave Telegram for Slack, a CRM or a script via send anywhere. If you are not sure Reddit is even the right channel, run the free report before you subscribe to anyone, including us.

What this playbook refuses to promise

If you want the same job inside a different product shape, the comparisons are Brand24, ReddGrow and ReplyGuy, listed from the alternatives hub and from compare. The older ones, if your current tool is a research tab or an email alert: GummySearch, F5Bot, Linkeddit, EaseClaw.

AI answers playbook: FAQ

What is an AI answer citation, in practice?

You ask Perplexity or Gemini for the best tool in your category and your product is named, usually because a public conversation already talked about it in a specific way. This playbook does not promise that ChatGPT will name you from Reddit. On 14 August 2026 Reddit's share of ChatGPT Search citations fell from 3.83% to 0.52%.

How is this different from the ChatGPT guide and the GEO hub?

The GEO hub and /geo/chatgpt are the map of how an answer engine picks a source. The get-cited-by-ChatGPT guide is the long page about your own site after the Reddit drop. This one is the weekly routine for Perplexity, Gemini and other AI answers: prompts, thread types, comment shape, account survival, and the Scale check.

Will Karmy get me cited by ChatGPT from Reddit?

No. We do not promise that. Karmy gets you into the threads and writes a first draft that is trying to be useful, not loud. On Scale, a weekly check of ChatGPT, Perplexity and Gemini tells you whether they currently name you. It does not make ChatGPT cite Reddit again.

How many comments a week?

On an account that has finished warming up, two careful comments in threads that match your three prompts beat a pile of keyword replies. On a new account, zero promotional comments. Follow the warm-up.

What does the routine cost?

Ten days free, no card, then €9 a month for one mission, €19 for up to five, or €39 on Scale for up to 20 missions. The weekly check of ChatGPT, Perplexity and Gemini is on Scale. A karmito is coming soon.

Run the routine from Telegram

Tell Karmy the product. The threads that match the question show up with a draft. You keep the part that makes it true. Free for 10 days, no card to start.

Join the waitlist →