The Answer Engine analytics platform built for agencies and brands. Get better brand visibility on ChatGPT, Google Gemini, Google AI mode. and more.
San Diego, CA
Joined January 2026
- Tweets213
- Following49
- Followers245
- Likes91
gethoverboard retweeted
Hoverboard AI tracked 6,785,895 citations across AI engines in September.
Only 22% pointed to the website of a brand mentioned in the answer.
The other 78% went to directories, review sites, news outlets, forums, blogs and listicles.
gethoverboard retweeted
Noticing some changes in how ChatGPT fans-out. It no longer does it all at once. It searches in rounds, and each round is written after reading the previous one. Here's an example, using GPT-6 Astra that was set on high effort...
There's often a huge difference in the brands that ChatGPT mentions, cites and recommends for paid users vs free users (due to the retrieval process).
A paid-user often has a higher model selected, and some degree of "thinking" (light, medium, high, etc.) which can result in far more fan-out queries and URLs retrieved.
With a free account (which is 90%+ of all ChatGPT users) there are usually significantly less fan outs and URLs retrieved.
In this example below, ChatGPT ran 28 fan-outs and retrieved 135 URLs for a single prompt ("What are the top luxury SUVs for 2027?") compared to just 3 fan outs and 33 URLs retrieved for a free, non-signed in user.
If you are optimizing for ChatGPT, your focus should be on the 90% of users.
ChatGPT visibility isn’t simply one number.
There are three layers to winning brand mentions, citations and recommendations on ChatGPT:
• What’s baked into the model's parametric memory
• What’s in ChatGPT’s own index
• What Google surfaces when it searches live
A brand may have strong parametric memory (what the model already knows) but may not "survive" the brand recommendation process after retrieval.
If you don’t know which layer is talking, you’re optimizing blind.
Post here: gethoverboard.com/the-three-…
Layer 2: First-party index visibility
Even if the model doesn’t “know” you, ChatGPT can still retrieve you from its own index (Labrador).
That’s crawl access, indexation, on-page relevance, and pages that match the questions people actually ask ( including fan-out queries).
Full breakdown here. How the 3 layers work, why answers vary so much, and what to do for each one:
gethoverboard.com/the-three-…
Layer 3: External-search visibility
This is the one marketers already understand: search engines (mainly Google).
When ChatGPT needs freshness or grounding, it searches the web. Strong SEO, reviews, comparisons, and third-party mentions still matter here.
ChatGPT has a pretty clear citation "authority ladder" as they try to decrease the number of domains with less trust and authority from impacting their responses.
Been tracking the URLs ChatGPT retrieves this week for a two hundred prompts in the health, legal, tech, shopping, food and a few miscellaneous prompts. Citations are getting tougher to come for smaller sites.
Out of the 6,938 retrieved urls, the average DR (ahrefs) was 53.3.
1,302 were used as sources (18.8%) and average DR was 62.8.
1,018 were citations (14.7%) and average DR was 65.8
For domains retrieved from parametric memory and used as fan-outs (we can see these with the use of "site:"), the DR for those was 75.6.
Oops! Made a mistake on the image. Should say 6,938 retrieved URLs analyzed (not 4,039). Looks like it's too late to edit.
Note: This was done using a combination of one signed in pro-account and a non-logged in free account. To avoid getting personalized responses that could have been impacted by previous conversations, unique topics and prompts were used.
Truth. They are all based on estimates, not actual data. The only source of real prompt data is the answer engines themselves.
There is no such thing as "prompt search volume." If you're paying for it, you're wasting your money.
I've actually looked at a bunch of data samples from clickstream providers who track what people type into ChatGPT.
Can you guess what I saw?
Multi-sentence paragraphs spanning multiple topics, dictated rambles with assignments, rather than searches and follow-up questions that make no sense without previous messages.
…and then there’s personalization, which further messes up the input that AI chatbot actually gets.
Bottomline:
Deriving any meaningful “search volumes” out of that mess is no better than guessing.
But is that data even important for tracking your AI visibility?
That’s the question I asked Dan Petrovic @dejanseo on the latest episode of @Ahrefs Podcast. Dan runs the reputable “AI Influence” agency called DEJAN AI and he is one of the most respected (and positively nerdy) figures in SEO, who’s been at the forefront of the industry for more than 15 years.
Here's what else we covered with Dan:
🔸 Why LLMs.txt is mostly pointless.
🔸 Why LLMs reject Reddit content over 90% of the time
🔸 What is “primary bias” and “secondary bias” of AI models
🔸 Which search indexes do ChatGPT, Claude, and Gemini rely on
🔸 The "Trojan horse method" for getting into LLM citations fast.
🔸 How Dan tests on-page changes with an agentic loop
🔸 What to actually track instead of prompt volumes.
🔸 How to generate the right prompts for AI visibility measurement.
Watch full episode here 👇
YouTube: youtu.be/roTPZBj4JSc
(also on Apple podcasts & Spotify)
Starting to see a lot of more 3650d "freshness" windows from ChatGPT fan-out queries, mostly tied to "site:" fan-outs. Mainly on health related prompts but have also seen it in legal, even on restaurant related queries so not just YMYL prompts. The only other domain I've seen it on is Reddit. They may not be using it much as a citation source, but clearly they still trust it and are using it to help with the answer response.
Almost always seeing this on high authority domains (average DR of 83.2 ). Clearly, they would rather retrieve URLs with content on it up to 10 years old over lower quality domains.
Recency window is one of the fields in ChatGPT's pipe-grammar search calls. It's the model telling its search backend, per query, "prefer results no older than N days." So in these examples, it means the model asked for FDA label pages with a ~10-year window. I have seen pages retrieved with creation dates outside of those windows, so it may be looking for pages modified within those windows, not necessarily created. So far I've seen 1, 2, 7, 30, 90, 365, 1000, 2000 and 3650d.
This is a small sample size here where I've seen this now in 100 prompt responses (and several hundred fan-outs), but clearly ChatGPT is really focused on retrieving URLs from high authority sites.
Hat tip to @suganthan, who pointed out the freshness windows and did an amazing write up of it last week. If you missed it, make sure to read his post (link in the comments below).
Suganthan's post: suganthan.com/blog/chatgpt-n…
Everyone by now has seen how Reddit citations have nose-dived in ChatGPT in August. We've seen the same thing in Hoverboard data. But the other models have been pretty steady. At least for the moment, Reddit is still important for Google's AI products.
Can we stop spamming Reddit now? 😂
We saw an 87% drop in Reddit / ChatGPT citations in August, with a gigantic drop on August 14th. ChatGPT 5.6 Luna (the default model for free accounts, which make up an estimated 90%+ of all users) was rolled out the week of August 6th. As the chart here shows, you can see big drops that week, before the hammer on August 14th.
Hoverboard AI analyzed 828,730 citations tracked in GetHoverboard.com over a two week period to get this data. If you compare it to July 14th, the drop is even greater (96%!).
Reddit went from about 1 in every 44 ChatGPT citations on July 14 to about 1 in every 1,000 after August 14.
Models can change the way LLMs answer and cite sources (as we can see here, sometimes drastically). This is why it's important to track AI visibility and citation sources on a consistent basis.
ChatGPT Reddit citations have taken a huge hit lately. We’ve noticed around an 87% drop, with citations falling off a cliff in August 14th.
🤖 Made with AI
We analyzed 143,295 ChatGPT responses and found byte-identical answers responses (days apart, citations included).
The duplicates are real, but we can’t prove caching caused them. Recently, @5eg and @resoneo ran an OpenAI experiment and matched prompts to precomputed answers.
Are we seeing it in the wild?
Full analysis, methodology and implications for AI visibility: gethoverboard.com/is-chatgpt…
