Two ways to show up in AI: remembered, or retrieved
Markup refreshed for current answer engines.
On 16 December 2024, ChatGPT Search opened to all logged-in users — the most-used chatbot in the world began searching the live web and citing its sources with links. Perplexity and Google had pointed this way already, but now it is mainstream, and it splits AI visibility cleanly in two.
the short answer
As of December 2024, there are two ways to appear in AI. The first is to be remembered — named from the model’s training, with no link. The second is to be retrieved — pulled into the answer by a live web search and cited with a clickable link, as ChatGPT Search now does for all logged-in users. Retrieval rewards content that is current, clear and relevant, published by an entity the search layer recognises, and it leaves a measurable trace: the outbound links carry UTM tags you can read in your analytics.
key takeaways
- On 16 December 2024 ChatGPT Search opened to all logged-in users (it had been Plus/Team only since 31 October) — the most-used chatbot began searching the web and citing sources.
- There are now two ways to appear in AI: remembered, from the model’s training (no link), or retrieved, from a live search (with a link to your page).
- Retrieval is a new, measurable channel: ChatGPT’s outbound links carry UTM parameters (chatgpt.com), so the visits land in your analytics as referral traffic.
- Being retrieved depends on being retrievable: content that is current and clearly relevant, and an entity the search layer recognises and trusts.
- This does not replace SEO or training-based authority — it adds a surface where recency, clarity and entity decide whether you are one of the few sources cited.
two paths to the same answer
Both rows start with the same question and end with an answer. The difference is the middle. The top path works from memory — what the model learned in training — and hands back an answer with nothing to click. The bottom path searches the live web, reaches a current page, and returns an answer with a link to it. Before December 2024 the bottom path was a niche experience; after it, it is what most people get by default. For a brand, that flips a quiet assumption: the answer a customer reads about your category may now be assembled from pages published this week, not from what a model absorbed a year ago, and the page it pulls might be one you published only yesterday.
Why the split matters for visibility
For years, "showing up in AI" meant one thing: being known to the model — absorbed into its training, named when relevant, invisible to measurement and slow to influence. That is still real and still worth earning. But treating it as the whole game now misses half the surface. The retrieved path behaves differently enough that it needs its own attention: it rewards freshness in a way training never did, it can lift a brand-new page into an answer within hours of publication, and it produces a link a real person can follow to you. A brand that has optimised only for being remembered — broad presence, strong associations — may still be passed over in a retrieved answer if its actual pages are stale, buried, or unclear about who wrote them.
The practical consequence is that the two paths deserve different work. Being remembered is earned slowly, through years of consistent, distributed presence until the model simply knows you. Being retrieved is earned page by page, by keeping the answer to each real question current, clear, and unmistakably yours. The strongest position holds both at once — known to the model and retrievable on demand — and December 2024 is the moment the second of those stopped being optional, because the chatbot most of your audience opens now searches before it answers.
Retrieval and citation, in three parts
The two ways to appear, what makes a page retrievable, and the one measurable upside the retrieved path brings. Open each layer for the part that changes how you earn AI visibility.
01 Remembered vs retrieved: two ways to appear
Until late 2024, a chatbot answering a question about your category was working from memory: whatever it had absorbed from its training data, frozen at the model’s cutoff. If it named you, that was a "remembered" mention — no link, no live source, just the model’s internalised sense that you belong in that answer. Retrieval adds a second path. When ChatGPT Search runs, the model does not rely on memory alone; it searches the live web, reads the current top sources, and writes an answer that cites them with clickable links. The two paths reward different things. Remembered presence is the slow accumulation of being everywhere, consistently, over time, until the model simply knows you. Retrieved presence is the immediate reward of having a page that is current, clear, and relevant enough to be pulled into the answer the moment someone asks. After December 2024, both paths are live for the audience that matters, because the chatbot most people open now does both.
02 Why retrieval rewards clarity and recency
A live search is, in effect, a fast and literal reader. It reaches for pages that answer the question directly and near the top, that are structured so a clean passage can be lifted out, and that carry enough credibility signals to be worth citing. This shifts the emphasis compared with being remembered. Recency starts to count in its own right: retrieval is trying to find the current answer, so a page that is obviously maintained and up to date has an advantage over one that has sat untouched, even if the older page is more authoritative in the abstract. Clarity of entity counts too — when the system can tell without ambiguity who published a page and what they are known for, it can trust the source enough to put it in front of a user with a link. None of these are loopholes. They are the ordinary qualities of a genuinely good answer, now being assessed by a machine that searches before it speaks, which is why the work that makes you retrievable is the same work that makes you worth retrieving.
03 The measurable upside: links you can see
The retrieved path comes with something the remembered path never offered: a trace you can measure. When ChatGPT Search cites your page, it does so with a real outbound link, and those links carry UTM parameters naming chatgpt.com as the source. That means a visit arriving from a cited answer shows up in your analytics as identifiable referral traffic — not a guess, an actual logged session you can attribute to this channel. It is an incomplete signal, because many people read the cited answer and never click, and that influence leaves no footprint. But the clicks that do come are countable, which is rare in this space; most AI visibility has to be inferred from sampling rather than read off a referral report. Recognising the chatgpt.com referral in your analytics turns retrieval from an abstract aspiration into something you can watch respond to the content you publish, which is exactly the kind of honest, observable feedback loop the rest of AI measurement mostly lacks.
Entity is the thread through both paths
Whether a model remembers you or retrieves you, the same quiet factor runs underneath: how clearly the system understands who you are. To be remembered well, your brand has to be a stable, consistent entity across the web — described the same way everywhere, associated reliably with your category — so the model forms a confident internal picture rather than a fuzzy one. To be retrieved and cited, a live search has to be able to look at a page and tell, without guessing, that it was published by a recognised source that is an authority on the question. In both cases the obstacle is ambiguity. A brand the system is unsure about gets remembered weakly and retrieved cautiously, because uncertainty makes a model hold back rather than risk a wrong attribution.
That is why entity clarity is the work that pays off on both paths at once. Making your identity unmistakable — consistent naming, clear authorship, the kind of structured signals that say plainly who published this and what they are known for — raises your standing in the model’s memory and your odds of being the page a search trusts enough to cite. It is not two separate projects. The discipline of being a clear, credible, well-defined entity is what lets the AC Group earn citations on both surfaces, and it is the same discipline we have built around for 27 years: be unmistakably who you are, and be genuinely the best answer.
AI that searches: quick answers
Is this the same as Perplexity or Google AI Overviews?
It is the same idea arriving on a much bigger stage. Perplexity built its whole product around retrieving live sources and citing them, and Google had been folding AI Overviews into search results through 2024; both already showed that an answer engine could read the web and attribute what it used. What changed in December 2024 is that ChatGPT — by a wide margin the most-used chatbot — opened the same behaviour to all of its logged-in users. So the mechanism is not new, but the audience is: a very large number of people who previously got answers only from the model’s training can now get answers assembled from the live web, with links to the sources. For anyone thinking about visibility, that turns "being a source an answer engine retrieves" from a niche concern about Perplexity into a mainstream one, because the surface that does it is now the one most people open by default.
Does retrieval replace training-based mentions?
No — it sits alongside them, and the two reward overlapping but distinct things. Being "remembered" means the model has absorbed your brand into its parameters from training data, so it can name you without searching; that depends on broad, consistent presence across the web over time. Being "retrieved" means a live search pulls your page into the answer right now and links to it; that depends on having content that is current, clearly relevant to the query, and recognisable as the work of a known entity. A strong position has both: the model knows you well enough to mention you unprompted, and your pages are fresh and clear enough to be pulled in when it searches. Neither replaces the other, and chasing only one leaves visibility on the table — the training-based mention gives you presence even without a search, while the retrieved citation gives you a live, clickable link the moment the topic is current.
How do I become a source ChatGPT Search retrieves?
By being retrievable, which is more concrete than it sounds. A live search reads the web the way a careful reader would, so the pages it tends to pull are the ones that answer the question directly and early, are structured clearly enough to extract a clean passage from, and carry signals that the source is credible and is who it claims to be. Currency matters more here than in training-based mention, because retrieval is reaching for the up-to-date answer; a page that is obviously current and maintained has an edge over one that has not been touched in years. Entity clarity matters too: when the system can tell unambiguously who published the page and what they are an authority on, it can trust the source enough to cite it. None of this is a trick — it is the same discipline of being genuinely the clearest, most credible answer, now read by a machine that searches before it speaks.
Can I measure traffic from it?
Partly, and better than from training-based mention, which is the quietly useful part. When ChatGPT Search links to your page, the outbound link carries UTM parameters identifying chatgpt.com as the source, so visits that come through show up in your analytics as referral traffic you can actually see and count. That is a real change: a citation that produces a measurable click, rather than an appearance inside an answer that leaves no trace. It is not a complete picture — plenty of people read the cited answer without clicking, and that influence is invisible to your analytics — but the clicks that do come through are attributable, which means you can watch this channel grow or shrink with the content changes you make. Set up to recognise the chatgpt.com referral and you have one of the few honest, countable signals in a measurement landscape that is otherwise mostly inference.
A note on sources and timing
This is written in December 2024, in the days after ChatGPT Search opened to all logged-in users on the sixteenth — the rollout that took retrieval-with-citations from a Perplexity-and-Overviews novelty to a default behaviour of the most-used chatbot. We have described what was observable then: the live web search, the cited sources in a side panel, the outbound links carrying chatgpt.com UTM tags. We have not reached for figures that did not yet exist — how often people click a cited answer, what share of traffic this becomes — because those were still unmeasured as we wrote, and borrowing later numbers backward would lend a false precision. What is solid is the structural shift: from one way of appearing in AI to two, remembered and retrieved, each earned differently and both now live for the audience that matters.