- A citation is a single-instance event, downstream of one retrieval. A recommendation (your brand named alongside competitors in a comparison or best-of answer) is a more durable, competitively-loaded state, and the two don't move together.
- Whether "being chosen" is actually stable depends far more on query intent than on the brand: a 14,000-API-call Conductor study found brand overlap between repeated runs ranging from 40% (Purchase intent) to 63% (Comparison intent), and lead-brand retention from 59% to 91% across the same range.
- The GEO paper's own simulation of simultaneous competitor optimization found the effect is redistributive, not additive. A tactic every competitor can apply at once closes gaps between brands rather than protecting whoever's currently ahead.
- What compounds instead is the off-site, earned layer (reviews, comparisons, press) which is slower to build and slower for a competitor to copy, unlike a content-structure change anyone can replicate the same week.
Mention vs. citation vs. recommendation already separates these into three outcomes that don't move together. This piece goes one level further into the third one: recommendation is the outcome most people actually mean when they say they want to "be chosen" by AI search, and the evidence on how stable that actually is, once you have it, is more specific and more conditional than "get recommended and you're set."
What does "being chosen" actually require beyond a citation?
A citation can happen with zero competitive framing: a single-source factual answer naming you because your page was the relevant source for one specific claim. A recommendation is different by definition: your brand appears in a list of options a model surfaces for a comparison or "best of" query, positioned against named alternatives it could have named instead. Every recommendation involves a mention. Most citations never involve a recommendation at all.
That distinction matters because the two outcomes respond to different levers. Content extractability (structuring your own pages so a crawler can lift a clean answer from them) raises your odds of being cited when your page is the relevant source. It does comparatively little to raise your odds of being recommended in a competitive comparison query, because that outcome depends on third-party proof the model can weigh against competitors, not on how well one page is structured.
Is being chosen actually stable, or does it churn like citations do?
This is the part worth checking rather than assuming. Conductor ran 14,000 API calls across 10 industries, seven intent types, and four large language models (ChatGPT, Perplexity, Claude, and Gemini) with 50 runs per industry/intent/engine combination, generating 1,225 unique pairwise comparisons per group. Their finding: intent type predicts AI brand-recommendation consistency more than the industry being asked about.
The consistency ranking, by brand overlap between repeated runs of the same prompt, from least to most stable:
| Intent type | Brand overlap between runs |
|---|---|
| Purchase | 40% |
| Support | 47% |
| Navigation | 47% |
| Recommendations | 49% |
| Pricing | 58% |
| Education | 60% |
| Comparison | 63% |
Out of every 10 unique brands that appeared across any two Purchase-intent prompt runs, only four showed up in both, close to a coin flip at the exact moment someone's actually about to buy. Lead-brand retention specifically, whether the top-named brand stays on top across repeated runs, ranged from 59% at Purchase intent up to 91% at Comparison intent. Comparison-style queries ("X vs. Y," "alternatives to X") produce a far more stable answer than a direct purchase-intent query does, for reasons the study ties to intent structure rather than any specific brand's actions.
The practical read: "are we the recommended brand" isn't one stable fact to check once. It's a different reliability question depending on which stage of a buyer's actual search behavior you're asking about, and the queries closest to an actual purchase decision are the least stable of all.
What does GEO-style optimization protect, and what doesn't it?
Even a brand that earns a strong recommendation position today isn't holding a fixed asset. The GEO paper tested what happens when every competing source applies the same optimization method at once, not just one source in isolation, and found the effect is redistributive rather than additive. In that simulation, a source originally ranked 5th gained roughly 115% relative visibility using the Cite Sources method, while the source originally ranked 1st lost roughly 30% of its visibility in the same scenario. The gap between them closed. The total pie didn't grow evenly for everyone.
That's the mechanism worth internalizing before treating any single content tactic as a durable moat: if a tactic is cheap enough that every competitor in a comparison set can copy it within a news cycle, it was never going to protect a lead. It was only ever going to reset the field closer to even.
What actually compounds?
The layer that doesn't reset the same way is the off-site, earned one (reviews, independent comparison content, genuine press coverage). Earned media accounts for 82-89% of AI citations across every disclosed-methodology study that's measured it, and unlike a content-structure change, a competitor can't simply copy your review volume or your press history by editing a page. It's slower to build for exactly that reason, and slower for someone else to erase once you have it, which is the actual shape of a durable advantage, as distinct from a one-time lift that closes the moment everyone else notices what worked.
Frequently asked questions
What's the difference between being cited by an AI model and being recommended by it?
A citation is a model naming or linking your domain as the source behind one specific claim. It can happen with zero competitive framing. A recommendation is a model surfacing your brand in a list of options for a comparison or best-of query, positioned against named alternatives. Every recommendation involves your brand being named, but plenty of citations happen with no competitive framing at all. See mention vs. citation vs. recommendation for the full breakdown.
Is "being the recommended brand" a stable, durable state, or does it churn like citations do?
It depends heavily on what the person was asking, not on the brand itself. A Conductor study of 14,000 API calls across 10 industries, 7 intent types, and 4 LLMs (ChatGPT, Perplexity, Claude, Gemini) found brand overlap between repeated runs of the same prompt ranging from 40% for Purchase-intent queries up to 63% for Comparison-intent queries. Lead-brand retention specifically ranged from 59% (Purchase) to 91% (Comparison): being chosen is far more stable for a comparison-style query than at the actual purchase moment.
Does GEO-style content optimization make being chosen more durable?
Not on its own, per the GEO paper's own simulation of every competitor optimizing at once. The effect was redistributive, not additive: a source originally ranked 5th gained roughly 115% relative visibility using the same method that cost the 1st-ranked source roughly 30% of its visibility, closing the gap between them rather than protecting an existing lead. A tactic every competitor can apply simultaneously isn't a durable advantage by itself.
So what actually builds durable "being chosen" status?
The same off-site, earned proof that drives recommendation in the first place (reviews, independent comparisons, press) rather than a one-time content or structure change. Earned media accounts for 82-89% of AI citations across every disclosed-methodology study that's measured it, and that's the layer that's slower to build and slower for a competitor to erase, which is exactly why it compounds instead of resetting every time someone else copies the same on-site tactic.
Sources referenced
- Conductor: AI Brand Recommendation Study — Why Intent Type Predicts AI Output Consistency (14,000 API calls, 10 industries, 7 intent types, 4 LLMs)
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, "GEO: Generative Engine Optimization" (arXiv 2311.09735, KDD 2024)
This piece extends mention vs. citation vs. recommendation and builds on the redistribution finding in the GEO paper explained. For why earned media is the layer that actually compounds, see why earned media outperforms owned content in AI citations.
