Reddit lost 86% of its ChatGPT citations in four days. Here is what that tells you about your own AI visibility

The one that actually matters: Reddit fell off a cliff inside ChatGPT

For two years the whole AI search industry has repeated the same line. Reddit is the kingmaker. Get quoted on Reddit, get quoted by the machine. We have said it ourselves. It is truth number three in our own playbook.

This week that line took a hammer blow.

Promptwatch tracking data, reported by Search Engine Land, shows Reddit held a steady 3.8 percent share of ChatGPT Search citations from 18 July through 7 August 2026. On 14 August that share collapsed to under 1 percent. The average across 14 to 17 August was 0.5 percent. That is an 86.4 percent relative drop in four days.

There were two separate slides, not one. The first started on 8 August, the same day Promptwatch says ChatGPT Search changed its query fan-out behaviour. The second, much bigger one started on 14 August. Nobody at OpenAI announced either. Publishers found out the way publishers always find out: by watching a number fall.

Be careful with this one. Search Engine Journal has pushed back and pointed out the drop is not fully explained, and that a data collection issue on the tracking side cannot be ruled out. This is one tracker, not a confirmed OpenAI policy change. Treat it as a very loud signal, not gospel.

Here is the part most people are missing. Google did not do the same thing. AI Overviews and AI Mode showed slower, milder declines for Reddit content over the same window. So this is an OpenAI-specific retrieval decision, not the internet deciding Reddit is worthless.

What this actually means for you

If your entire AI visibility plan was "get mentioned in a Reddit thread and pray", you just watched your plan get benched by a config change you were never told about. That is not a Reddit problem. That is a single-point-of-failure problem.

The operators who did not blink this week are the ones with their brand named across five or six independent sources: a listicle, a directory, a YouTube description, a LinkedIn profile, a local press mention, their own answer-first pages. When one source gets down-weighted, the other five carry them.

We call this source-stacking, and we walk through the exact stack, in order, inside the Dojo. It is not clever. It is just deliberate.

Gemini 3.7 Flash is now answering your customers in AI Mode

Late on Friday 14 August, Google's Robby Stein and Rajan Patel confirmed on X that Google Search is now running Gemini 3.7 Flash on some AI Mode queries. The model had been announced the day before, but Google said nothing about Search at the time. It was billed as its "most intelligent workhorse model yet for coding and agents".

Translation: the thing summarising your industry to your customers got swapped out on a Friday night and you were not consulted.

This is the rhythm now. Model changes ship without warning, citation patterns shift, and the people who track their AI visibility notice within days while everyone else notices in a quarterly report. If you are not measuring, you are not in the game. We built a free way to do it and wrote it up in our AI visibility tracking guide.

Google will let you switch yourself off in AI answers. Almost nobody should

Search Console now carries a control under Settings, called Search generative AI, that lets a site owner opt their content out of AI Overviews, AI Mode and AI Overviews inside Discover. Google began testing it on 3 June 2026 under pressure from the UK Competition and Markets Authority, and it took effect on 17 June. It started with a subset of UK operators and has been widening since.

Read the small print, because this is where people are going to hurt themselves:

• Opting out removes your traffic and impressions from AI Overviews, AI Mode and AI Overviews in Discover.
• It does not change your normal ranking. Flipping the switch will not push you up or down in the blue links.
• It does not govern whether your content is used to train Google's models. Different question, different lever.
• Changes take one to two days to propagate.

So who should press it? A publisher whose entire business is ad impressions on their own pages, who has done the maths and knows AI referrals are worth less than the cannibalisation. That is roughly it.

If you are a plumber, a law firm, a SaaS, a local trade, an ecommerce brand, anything where the goal is to be recommended: pressing that button is voluntarily deleting yourself from the fastest-growing recommendation surface in the world. Do not do it because a LinkedIn post made you angry about AI.

AI Mode is now reading your customer's inbox

Google has been rolling out Personal Intelligence in AI Mode, which lets users connect Gmail and Google Photos so answers can draw on their own context: flight confirmations, hotel bookings, past purchases, photos. It runs on Gemini 3, it is opt in, and it has been going out to AI Pro and AI Ultra subscribers on personal accounts, not Workspace business, enterprise or education accounts.

Think about what that does to a query. "Best waterproof jacket" stops being a generic question. It becomes "best waterproof jacket for someone flying to Reykjavik on the 4th who already owns two North Face shells". The answer gets narrower. Fewer brands make the cut.

The businesses that survive that squeeze are the ones the model can describe precisely: one clear name, one clear job, one clear service area, said identically everywhere. Vague brands get filtered out first. That is entity clarity, and it is the least glamorous, highest-return job in AI search. The full walkthrough is in the AEO playbook.

Apple quietly added 4,656 crawler IPs

Apple has added over 4,000 new IP addresses for Applebot, the crawler that feeds Siri and Apple Intelligence. Specifically 18 new /24 ranges and 3 /28 ranges, which works out at 4,656 addresses.

Nobody wrote a think piece about this. They should have. You do not multiply your crawl capacity by that much unless you intend to crawl a great deal more of the web, a great deal more often. Apple has hundreds of millions of devices with a default assistant and, until now, a fairly thin index behind it.

Two things to do about it, both boring, both five minutes: check your robots.txt is not blocking Applebot by accident, and check your server or WAF is not rate-limiting a crawler you have never heard of. We have seen sites lock out AI crawlers by accident and then wonder why they are never quoted. Free tools for checking this live in the Armory.

The llms.txt reality check nobody wanted

A crawler census published by DEJAN and picked up by Search Engine Roundtable this week looked at what the big crawlers actually requested on their own domain over 30 days. Google, Anthropic and OpenAI crawlers fetched robots.txt 1,150 times. They fetched llms.txt four times.

Four.

We said this in our llms.txt piece and we will say it again: it is a five minute nice-to-have, not a religion. Add it, move on, and spend the saved afternoon on something the machines are demonstrably reading. Anyone selling llms.txt as the switch that fixes your AI visibility is selling you a dream at $2k a month.

The rest of the tape

That last one deserves a sentence. Search Engine Roundtable spotted ChatGPT Ads swapping which advertiser is shown when the same answer is reloaded. That is inventory optimisation behaviour. It is what a maturing ad product does. The free-answer window in ChatGPT is not going to stay this uncommercial forever, which is exactly why earning the organic citation now is worth more than it will be in eighteen months.

What to actually do before Friday

1. Run the source audit. List every place on the open web that names your business. If Reddit is more than a third of that list, you are exposed. Fix the ratio, not the platform.
2. Re-baseline your prompts. Ask ChatGPT, Gemini and Perplexity the five questions your customers actually ask. Screenshot the answers. Date them. Model swaps mean last month's baseline is stale.
3. Check crawler access. robots.txt, WAF rules, rate limits. Applebot, GPTBot, ClaudeBot, Google-Extended. Confirm you are not blocking the machines you want quoting you.
4. Leave the AI opt out alone unless you have run the revenue maths and can show the number.
5. Pick one answer-first fix. One page, one question as a heading, a 40 to 60 word clean answer directly underneath. That is the single highest-leverage thing on this list and it takes twenty minutes.

Point five is the one agencies quietly charge $2,000 a month to do for you. It is an afternoon's work once someone shows you the order to do it in. That is literally what the Dojo is for.

What we got wrong, and what we did not

We have told you Reddit is the kingmaker. This week, in ChatGPT specifically, that stopped being true almost overnight. We are not going to pretend otherwise.

What did not change is the reason we said it. Reddit mattered because it was third-party, independent, and full of people describing businesses in plain language. Those properties still win citations. They just are not exclusive to one website. The tactic was fragile. The principle was not.

That is the difference between chasing hacks and building an asset. Hacks get deprecated on a Thursday night with no changelog. Assets do not. If you want the version of this that survives the next config change, we teach it step by step in how to get AI to recommend your business.

The 2026 view from here

The pattern of this year is now obvious. Retrieval is getting more selective, not less. Models are being swapped mid-flight. Personalisation is narrowing every answer to a shortlist. Ad layers are being built on top. Every one of those forces points the same way: fewer names get mentioned, and the ones that do get mentioned are the ones the machine can describe confidently from multiple independent sources.

You cannot control the config changes. You can control how many places on the internet say clearly who you are and what you do well. That is the whole job. Everything else is noise with a newsletter.

Want the full system, the templates, the prompt library, the tracker and a straight answer when something like the Reddit drop happens? Join The Dojo. $67/mo. Or if you would rather we just did it, look at Done For You.