In-Line Citations vs. Retrieved Sources: What They Are and How to Track Them

In-Line Citations vs. Retrieved Sources: What They Are and How to Track Them

September 3, 2026 • Trackerly

On 6 August 2026, OpenAI rolled GPT-5.6 out to every ChatGPT user. Within days, Reddit's share of ChatGPT's citations fell from 3.83% to 0.52% — an 86% drop in four days, reported by Search Engine Land and Axios.

Here's the part that got less attention. ChatGPT didn't stop reading Reddit. In our own tracking, roughly one in four of the pages ChatGPT pulled while researching an answer was still a Reddit thread right through the collapse. What changed was whether those threads made it into the answer.

Two numbers. One fell off a cliff, one didn't move. If you were tracking a single "sources" figure, you saw a blend of both and understood neither.

That's the problem this post is about, and the reason Trackerly now records (as of August 21st) every source an AI engine returns as one of two distinct things.

What AI search engines actually return

When a model answers with web search enabled, its response carries two structurally different sets of URLs.

In-line citations

An in-line citation is a URL the model tied to a specific passage of its answer. You see them as the source pills inside ChatGPT's text, the [1] [2] markers in Perplexity, the numbered chips beside a claim in a Google AI Overview.

The engine is making a claim about this URL: this page is why I said that sentence.

Retrieved sources

A retrieved source is a URL the engine pulled up during research and never anchored to anything. Depending on the engine these appear in a "sources" list at the end of the answer, in a side panel, or nowhere visible at all.

The page was relevant enough to open, but the url itself is not tied to any specific part of the text.

The two behave independently

The August shift is the clearest demonstration yet that these are separate things. Reddit's retrieval held steady while its citation collapsed — the model kept consulting the same types of pages but stopped explicitly crediting them at the previous rate.

Why only one of the two makes you visible

Being retrieved feels like progress, but recent research suggests it mostly isn't — and that the ladder has more rungs than most people track.

Retrieval and citation are different stages, and content changes only affect the second. A July 2026 critical survey by Olivier Martinez reviewed 45 GEO studies published between November 2023 and July 2026. Its conclusion on what actually works: on-page techniques "can causally alter [already-retrieved content's] citation or use, but no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability." The widely quoted visibility gains from the original GEO paper, it notes, are "conditional on a source already being present in a fixed context." In other words: everything you can do to your page operates on the retrieved-to-cited step. Nothing reliably operates on the step before it.

A citation is not a mention. Semrush's June 2026 "ghost citations" study analysed 3,981 domain appearances across 115 prompts, 14 countries and four AI platforms and found that 61.7% of citations don't result in the brand being named in the answer at all. The engine links your page as a source while the reader never sees your name.

A citation is not a recommendation either. Lily Ray ran 100 B2B "best software" queries through Google AI Overviews across April, May and June 2026. Of 323 times a brand's own "best of" listicle was cited, Google left that brand out of its actual recommendation 224 times — 69% — frequently recommending the competitors named inside the cited article instead. Her summary: "a citation is not a recommendation."

The click, when it happens, is worth having. First Page Sage analysed 3.4 billion sessions across 218 client sites from January 2023 to July 2026. AI platforms grew from 0.1% of sessions to 6.2%, and ChatGPT visitors converted to leads at 4.7% against 1.9% for Google organic. That click only exists for a cited source. A retrieved one has no link to click.

So the ladder is: retrieved → cited in-line → named → recommended. Most tools report the bottom rung and the second rung as one number. The rest of this post is about separating them.

The engines don't behave alike

The volume of each tier varies enormously by engine, which is the main reason a blended source count can't be compared across them.

Orbit Media's study, published 2 September 2026 from 13,184 citations across 1,765 answers collected through 23 August, found the number of cited sources per answer ranged from 19.2 on Perplexity to 8.1 on Gemini, 4.5 on ChatGPT and 3.6 on Claude. Claude never cited Reddit once across hundreds of B2B answers.

GPT-5.6 in particular pulled the two tiers further apart. DecaGEO measured GPT-5.4 and GPT-5.6 against the same fixed prompts over consecutive weeks in July. GPT-5.6 nearly doubled the searches it ran per response — 5.8 to 11 — while halving the citations it produced, 12.1 to 6.2, and scoped around 90% of those searches to specific domains with the site: operator, up from 34–38%. More reading, less crediting.

And what it does credit shifted toward a specific kind of page. Something Inc summarised Promptwatch and Petra Labs data on what replaced Reddit: help centres and product documentation rose from around 2% of ChatGPT citations to 32%, established company sites from 4% to 18%, app marketplaces from 2% to 17%, while smaller sites and blogs fell from 66% to 32%. Their line: "Every category that gained is a first-party, structured, maintained surface with a clear owner."

What we are seeing

Here's the same question asked of Trackerly's own data over the last two weeks:

Engine Cited % of Total Retrieved URLs
ChatGPT 14.7%
Perplexity 53.2%
Google AI Mode 86.9%
Google AI Overviews 91.9%
Claude 48.7%

ChatGPT is the most source-hungry engine by a distance, retrieving roughly six retrieved sources for every one it cites. Google's two surfaces run the other way entirely, citing around nine of every ten URLs they surface, and between them produce 65% of all in-line citations in the data on 38% of the sources.

Which formats actually get cited

Format Cited % of Total Retrieved URLs
Video platform 85.8%
Social media 80.9%
Listicle / roundup 70.4%
Blog / editorial 69.6%
Ecom / product page 63.8%
Review 13,957
Brand website 62.5%
News / web article 54.4%
Directory / listing 48.7%
Press release 33.2%
Government / official 26.8%
Documentation / wiki 25.5%
Academic / research 21.3%
Forum / discussion 10.2%

News is over half of everything returned and lands almost exactly on the average. The edges are where the lesson is.

At the top sit formats that make a specific, liftable claim. A review says a thing is good. A roundup ranks it. A product page states a price. The model can point at one sentence and move on.

At the bottom, and this is the counterintuitive part: academic papers and government pages are among the least-cited formats we measured, at 21% and 27% against a 52% average. Models read authoritative reference material and then cite a blog post — not because the blog is more trustworthy, but because it already did the work of turning the reference into a quotable sentence.

Being authoritative gets you retrieved. Being quotable gets you cited. And per the August redistribution, being a maintained first-party page with a clear owner increasingly gets you cited by ChatGPT specifically.

What's new in Trackerly

The Citations tab is now Citations & Sources, with a scope selector under the heading:

  • All — every URL an engine returned. Unchanged from what you saw before.
  • In-line Citations — only URLs anchored to a passage of the answer.
  • Retrieved Sources — URLs pulled during research but never anchored.

The selector drives the whole section: the domain list, the over-time chart, the format breakdown and the trend deltas.

Individual sources now carry an "In-line" badge in the response detail view and on public prompt pages, so you can open a single answer and see exactly which URLs it leaned on.

The API exposes both, opt-in, on the responses and report endpoints:

  • citation_scope=all|cited|retrieved — narrows every citation aggregate, including both trending windows
  • include_citation_tiers=true — adds citation_type to each source and a summary.by_tier block to reports

Requests that send neither parameter return exactly what they did before, so existing integrations are untouched.

One caveat on history: tier labelling went live on 21 August 2026.

What to do with it

Measure cited share in addition to simply source count. Your source count going up is ambiguous. Your in-line count going up is not.

Set the benchmark per engine. A 15% cited rate is unremarkable on ChatGPT and alarming on Google AI Overviews.

Diagnose high-retrieval, low-citation pages as a content problem, not a visibility one. If a model opens your page constantly and never anchors to it, relevance is already settled — it chose to read you. What's missing is a liftable claim. That's the step the research says on-page changes actually move.

Compare engines inside one tier. ChatGPT will always look the most source-heavy under All, because it reports everything it opened.

Watch for tier drift — it's the early warning. A domain sliding from cited to retrieved is a downgrade: the model still finds you, it just stopped vouching for you. Reddit's August shift was exactly this, and in a combined count it was nearly invisible.

FAQ

What is the difference between an in-line citation and a retrieved source? An in-line citation is a URL the AI model anchored to a specific passage of its answer — it is asserting that the page supports that sentence. A retrieved source is a URL the model opened during research but never attributed anything to. Both come back in the same response; most tools count them together.

Is a retrieved source worthless? No. Retrieval is the prerequisite — nothing gets cited that wasn't retrieved first — and retrieved sources sometimes appear in a visible sources list. But the research is consistent that uncited sources get little exposure, no click, and no mention. Treat retrieval as making the consideration set and in-line citation as winning it.

Why is ChatGPT's cited percentage so much lower than Google's? Because they report different things. ChatGPT hands back every URL it opened while researching; Google AI Overviews publish something closer to a finished bibliography. Neither is better — it means you can't compare the two without splitting the tiers first.

Did GPT-5.6 change this? It widened the gap. Independent measurement shows GPT-5.6 roughly doubled the searches it runs per answer and halved the citations it produces compared with GPT-5.4, with citation share moving toward documentation, help centres and established first-party sites. Our own data, collected entirely after the rollout, shows ChatGPT returning about six retrieved sources per in-line citation.

Does every AI model support this? Every search-enabled engine we track does. Models answering from training data alone return no sources at all, so there's nothing to tier.

Did this change my historical numbers? No. The All scope is identical to what the section showed before. Tiers are additive.

Sources

Understand how AI is talking about your brand

Track how different AI models respond to your prompts. Compare OpenAI and Google Gemini responses to increase your visibility in LLMs.

Start monitoring AI responses →