Something Inc.Schedule a free consultation
GEO

Parametric authority vs live retrieval: where your GEO budget belongs

One clickstream study found ChatGPT triggering web search in anywhere from 15% to 66% of sessions depending on the model version. Whether your brand lives in the model's memory or in what it fetches decides which half of your GEO budget is doing anything.

TTTyler TruffiManaging Partner · AUG 12, 2026 · 10 MIN READ
15–66%
of ChatGPT sessions triggered web search, by model version
~100%
of Perplexity queries run a live web search
2
distinct memory systems your GEO plan has to serve
TL;DR · 60 SECONDSAI engines answer from two places: parametric memory frozen at training time, and retrieval pulled live from the web. Perplexity and Google's AI surfaces retrieve on essentially everything. ChatGPT, Claude, Copilot and the Gemini app decide per query, and one clickstream study saw ChatGPT's search rate swing between roughly 15% and 66% as models updated. Content structure fixes retrieval problems. Corroboration and consistency fix parametric ones. Funding only one leaves half your visibility untouched.

When an engine names a competitor instead of you, the useful first question is not what content to write. It is where that answer came from — the model's own trained memory, or a page it fetched thirty seconds ago. Those two failures look identical in a visibility report and need completely different work.

Duane Forrester laid the distinction out in Search Engine Journal in June: parametric memory is knowledge baked in during training and frozen until the next training cycle, while retrieval is fresh content pulled in real time when someone asks. Most GEO programmes are built as though only the second one exists, because the second one is the one you can affect this week.

Two memory systems, two different jobs

PARAMETRIC MEMORYLIVE RETRIEVAL
Where it comes fromModel training data, frozen until retrainWeb pages fetched at query time
How fast you can change itMonths to a training cycleDays to weeks
What fixes itConsistent, corroborated presence across sources the model will train onFindability, extractable structure, machine access
How it failsThe model 'knows' something outdated or wrong about youThe model cannot find or parse a page that answers the question
Where you see itConfident answers with no citationsCited answers naming somebody else's page

That last row is the cheapest diagnostic in the whole discipline. An answer delivered confidently with no sources attached is the model speaking from memory. An answer with three citations that are not you is a retrieval loss. You can sort most of your problem queries into those two buckets in an afternoon with no tooling at all.

Which engines retrieve and which decide

The engines split into two camps, and the split is stable enough to plan against.

Retrieval-dominant
Always-retrievePerplexity runs a live web search on essentially every question and shows its sources by design. Google's AI Overviews and AI Mode also lean on retrieval, but pull from Google's search index rather than Gemini's trained memory — which is why classic SEO transfers there more directly than anywhere else.
Mixed
Model-decidedChatGPT, Claude, Microsoft Copilot and the Gemini app judge per query whether to answer from parameters or go fetch. Claude invokes search as a tool when it decides it needs to; Copilot grounds against the web conditionally, with admin controls in the enterprise tier.

The planning consequence is straightforward. Retrieval-dominant surfaces respond to the work you can do now: clean structure, direct answers near the top, crawlable pages, coverage of the questions buyers actually ask. Model-decided surfaces respond to that work only on the fraction of queries where the model chose to search. On the rest, you are being described from memory, and memory was written months ago.

There is a second-order effect worth naming. Because Google's AI surfaces retrieve from the search index rather than from Gemini's trained memory, they are the one place where your existing SEO investment carries over almost intact. That makes them the cheapest AI surface to compete on and, for most B2B categories, the wrong place to judge your overall AI visibility — strong performance there can mask a parametric problem that only shows up when somebody asks ChatGPT the same question without triggering a search.

Which means a single blended AI visibility score is close to useless for budgeting. Two engines can disagree about you for structurally different reasons, and averaging them produces a number that points at no specific work. Reporting per engine is the minimum; reporting per engine and per memory layer is what actually tells you where the next pound goes. That separation is the same one behind how AI engines disagree on sources.

Why the retrieval rate moves under you

Here is the number that should reframe how you read your own tracking. A clickstream study of ChatGPT found web search triggering in somewhere between roughly 15% and 66% of sessions across the study window, with the rate shifting as the underlying models were updated.

High end of observed range66%
Low end of observed range15%

Observed range of ChatGPT sessions that triggered a web search, by model version (clickstream study)

A four-fold swing in how often an engine bothers to search is not something you can out-optimize. It is something you have to be positioned for in both directions.

Think about what that does to a visibility trend line. Your citation count on ChatGPT can halve in a month without a single change on your side, because a model update made the engine more willing to answer from memory. Teams reading that as a content failure will go and write more content, which addresses the retrieval layer that just got less influential. Understanding this is most of why AI visibility tracking through market share swings needs control metrics rather than raw counts.

Splitting a GEO budget across parametric and retrieval

The two layers do not need equal funding, and the right ratio depends on where your buyers actually ask. But both need a line.

1Retrieval work — the fast halfExtractable structure, direct answers high on the page, comparison and alternatives coverage, crawlability, clean machine access. Measurable in weeks, and it is the only half that responds to a quarterly content calendar.
2Parametric work — the slow halfConsistent descriptions of what you do across every surface a model will train on: your own site, Wikipedia and Wikidata where you qualify, G2 and Capterra, Reddit threads, podcast transcripts, conference bios. Corroboration across independent sources is the mechanism.
3The shared halfNamed authors with real credentials, cited sources, and factual consistency serve both layers at once. If budget is tight, this is where it goes first, because it is the only work that pays into both memory systems.

Parametric work is genuinely uncomfortable to fund because the feedback loop runs in training cycles rather than sprints. You will not see it land this quarter. What you can verify is the input: whether the description of your company is consistent everywhere a crawler can reach it, and whether independent sources corroborate the same facts. That consistency is what authority consolidation is actually for, and it is why scattered, contradictory positioning is more expensive in an AI-mediated market than it was in a search-mediated one.

Diagnosing which layer is failing you

1This weekSort your problem prompts by citation presence
THE MOVES
Run your top 30 buyer prompts across ChatGPT, Perplexity, Claude and Google AI Mode
Record whether each answer carried citations and whether any were yours
Bucket into: cited, not you (retrieval loss) and uncited, wrong about you (parametric loss)
DONE WHENEvery problem prompt sits in one of two buckets with a named owner.
2Same weekTest the memory directly
THE MOVES
Ask each engine to describe your company with web search explicitly disabled where the interface allows it
Record what it gets wrong, outdated or confused with a competitor
Treat every error as a corroboration gap rather than a content gap
DONE WHENYou know what the models believe about you without the web propping them up.
3Next monthFix retrieval losses with structure
THE MOVES
For each retrieval-loss prompt, check whether a page of yours actually answers that question directly and near the top
Verify the page renders without JavaScript and is reachable by AI crawlers
Add comparison and alternatives coverage where the losing prompts are decision-stage
DONE WHENEvery retrieval-loss prompt has a crawlable page that answers it plainly.
4Ongoing, quarterly reviewFix parametric losses with corroboration
THE MOVES
Standardise your one-line description and category across every third-party profile
Pursue coverage in sources with high training-corpus likelihood rather than chasing raw domain authority
Re-run the memory test each quarter and log what changed
DONE WHENIndependent sources describe you the same way, and the drift is tracked.

The second play is the one most teams have never run, and it is usually the most alarming. Turning search off and asking an engine to describe your company surfaces stale funding rounds, retired products, merged competitors and outright confusions — none of which any amount of new blog content will fix in the near term. It is also the clearest argument you will ever have for funding the slow half of the work, which is why we run it at the start of generative engine optimization engagements rather than at the end.

Where to start this quarter

Run the two diagnostics — the citation sort and the memory test — before you commission another piece of content. They take a day between them and they will tell you whether your visibility problem is something a content calendar can solve or something only consistency and corroboration will.

Then fund both lines, weighted toward whichever engines your buyers actually use. If your category lives on Perplexity and Google AI Mode, retrieval work carries most of the load. If your buyers ask ChatGPT and Claude open-ended questions and take the first confident answer, you have a memory problem, and the fix started six months before you noticed. The content that serves both is the content with named authors, real sources and facts that match everywhere else you appear.

DO THIS NEXTSort your top 30 buyer prompts into retrieval losses and parametric losses, run the memory test with search disabled, and put a separate budget line against each before commissioning more content.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.

TT
Tyler TruffiMANAGING PARTNER, SOMETHING INC.

Tyler leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.

Free consultation

Let us be the last SEO agency you ever work with

A 30 minute call and a free audit of your SEO and GEO position. You keep the findings either way.