Foglift
Buyer Intent Reshapes AI Citations: Discovery, Shortlist, and Variation Cite Different Webs
Across 375 buyer-intent responses from ChatGPT, Claude, Gemini, Google AI Overview, and Perplexity, when the buyer's intent shifts from discovery to shortlist to variation, the set of cited domains turns over almost completely. The average overlap across 25 verticals is 13.4%.
Last updated Data window 2026-05-18 (375 buyer-intent responses across 75 prompts × 5 engines)Download CSV
Methodology
Derived from the Q2 2026 AI Search Citation Benchmark (75 brand-neutral buyer-intent prompts × 25 verticals × 5 production AI search engines, 375 total responses, collected 2026-05-18 with each engine's grounded production model). For this analysis each vertical contains three prompts, one per buyer intent: discovery ("best X for Y"), shortlist ("X vs. Y vs. Z"), and variation ("cheapest X", "X for beginners", "X under $50"). For each (vertical, intent) cell we take the union of normalized root domains cited across all five engine responses, then compute Jaccard similarity between intent pairs within the vertical. Per-intent aggregator-vs-vendor breakdowns reuse the 15-category domain taxonomy from the parent benchmark, collapsed into the same aggregator-family vs. vendor-first-party lens used in the Day 2 artifact. The taxonomy, classifier, and aggregation script are reproducible from the files linked at the bottom.
The finding
Three buyer intents, three different webs. When the same vertical is queried with a discovery prompt, a shortlist prompt, and a variation prompt, the AI engines reach for almost entirely different domains for each one. Across 25 industries the average Jaccard overlap between any two intent pairs within a vertical is just 13.4%. The typical vertical produces a union of 51 distinct cited domains across its three intent cells, and only 2.2 of them (4.3%) appear under all three intents.
Most AI search optimization advice treats "the citations for our category" as one set. The data says it is three sets, with the boundary between them at the prompt itself.
Avg pair overlap
13.4%
Discovery ↔ Shortlist
18.9%
Discovery ↔ Variation
11.4%
Shortlist ↔ Variation
9.9%
Why it matters
Most teams treat AI search visibility as a single audit. They run one prompt set, see who AI engines cite, and decide on a content plan from that one snapshot. The Jaccard numbers above say that snapshot is capturing roughly a third of the citation surface at best, and roughly a tenth of it for variation-style queries.
The three intents map to three different funnel stages. Discovery is the buyer asking "what are my options." Shortlist is the buyer narrowing the field. Variation is the buyer testing the boundary cases. Each is a different question and AI engines pull from a different segment of the web for each one. A team that optimizes only for "best of" listicles wins the discovery surface and loses the other two.
By intent: where each one reaches
Each row collapses citations into three buckets: aggregator family (review aggregators like G2 and Capterra, niche publisher hubs like Sleep Foundation and Healthline, listicle content farms, lifestyle media, tech press, business press), vendor first-party (the brand's own canonical domain), and other (community UGC, video, marketplaces, institutional, dev platforms). Numbers are citation-level shares across 125 responses per intent (25 verticals × 5 engines).
| Intent | Aggregator | Vendor | Other | Citations |
|---|---|---|---|---|
| Discovery | 20.9% | 29.3% | 49.8% | 880 |
| Shortlist | 24.7% | 26.4% | 48.9% | 798 |
| Variation | 16.6% | 27.6% | 55.8% | 905 |
Three patterns sit inside the numbers. Shortlist is the peak aggregator intent at 24.7%, consistent with the comparison-page reality of "X vs. Y" queries: AI engines reach for the editorial outlet that already ranks the head-to-head. Variation is the peak "other" intent at 55.8%, driven mostly by community UGC and marketplaces. When a buyer asks for "cheapest" or "for beginners" or "alternative to," AI engines pull from Reddit threads, YouTube comparison videos, and Amazon listings far more often than they do for the other two intents. Discovery is the most vendor-friendly intent at 29.3% vendor share, but only by a narrow margin.
Domain personalities by intent
The top-cited domain across all three intents is the same: YouTube, showing up in roughly half of every vertical. After that the lists diverge sharply. The table below shows the top eight domains for each intent, ranked by how many of the 25 verticals cited that domain at least once.
Discovery
"Best X for Y use case" (open-ended exploration)
- 1.youtube.com13
- 2.forbes.com6
- 3.cnet.com5
- 4.medium.com4
- 5.salesforce.com4
- 6.techradar.com3
- 7.ventureharbour.com3
- 8.thesmarketers.com2
Shortlist
"X vs. Y vs. Z" (head-to-head comparison)
- 1.youtube.com12
- 2.forbes.com8
- 3.medium.com3
- 4.pcmag.com3
- 5.emailvendorselection.com3
- 6.healthline.com3
- 7.semrush.com2
- 8.consumerreports.org2
Variation
"cheapest X", "X for beginners", "X alternative" (qualified subset)
- 1.youtube.com11
- 2.reddit.com5
- 3.forbes.com4
- 4.zapier.com3
- 5.thegoodtrade.com3
- 6.emailvendorselection.com3
- 7.amazon.com3
- 8.outdoorgearlab.com2
Reddit shows up in the variation column but not the other two. Amazon and Apple show up only in variation. Forbes and CNET cluster in discovery and shortlist. PCMag and Consumer Reports anchor shortlist. emailvendorselection.com and ventureharbour.com (B2B listicle farms) show up in shortlist and variation. The "what site does AI cite for my category" answer is genuinely a different domain list per intent.
Intent-exclusive domains
The cleanest signal of intent personality is which domains show up under exactly one intent across two or more verticals and never under the other two. These are domains AI engines associate with a specific funnel stage.
Discovery only
- thesmarketers.com (2v)
- saashero.net (2v)
- larksuite.com (2v)
- dev.to (2v)
Shortlist only
- marketbetter.ai (2v)
- org.uk (2v)
Variation only
- amazon.com (3v)
- crmleaf.app (2v)
- martal.ca (2v)
- apple.com (2v)
- google.com (2v)
- prevention.com (2v)
By vertical: the most intent-fragmented categories
Averaging the three pair-Jaccards gives a single intent-coherence score per vertical. A vertical with a high average shares many cited domains across all three buyer intents (AI engines have a stable category canon). A vertical with a low average pulls from almost entirely different domain pools for each intent (no stable canon, the AI's answer depends heavily on the wording of the question). The 10 most intent-fragmented verticals, ranked low to high:
| # | Vertical | D ∩ S | D ∩ V | S ∩ V | Avg | Union | All-3 |
|---|---|---|---|---|---|---|---|
| 1 | Developer infrastructure / dev tools | 2.6% | 0.0% | 2.3% | 1.7% | 64 | 0 |
| 2 | AI tools / GEO / AI search optimization | 7.3% | 5.0% | 0.0% | 4.1% | 62 | 0 |
| 3 | Real estate platforms | 12.8% | 0.0% | 0.0% | 4.3% | 51 | 0 |
| 4 | Skincare & beauty products | 12.9% | 2.3% | 2.6% | 6.0% | 53 | 0 |
| 5 | Supplements & vitamins | 3.6% | 4.5% | 12.5% | 6.9% | 53 | 1 |
| 6 | Online learning / MOOCs | 17.5% | 2.0% | 2.4% | 7.3% | 62 | 1 |
| 7 | Sustainable / ethical apparel | 12.9% | 7.0% | 2.6% | 7.5% | 53 | 1 |
| 8 | Product analytics / web analytics | 18.5% | 2.4% | 4.8% | 8.6% | 51 | 0 |
| 9 | CRM software | 21.4% | 1.6% | 3.6% | 8.9% | 74 | 0 |
| 10 | Design / UI software | 10.9% | 12.0% | 6.8% | 9.9% | 69 | 2 |
Developer infrastructure leads at 1.7% average overlap, meaning the cited domains for "best dev tools for X" are nearly disjoint from "compare dev tool A vs. B" which is nearly disjoint from "open-source alternatives to Y." For a dev tools vendor the practical implication is that one piece of optimized content cannot rank for all three intents at once; each intent needs its own surface.
The five most stable verticals
At the other end, verticals where AI engines have a stable category canon across all three intents. These are categories with long-standing editorial authorities that get cited regardless of how the question is phrased.
| # | Vertical | D ∩ S | D ∩ V | S ∩ V | Avg | All-3 |
|---|---|---|---|---|---|---|
| 1 | Mattresses & bedding | 36.4% | 48.0% | 47.6% | 44.0% | 8 |
| 2 | Outdoor / camping gear | 25.0% | 28.6% | 17.9% | 23.8% | 4 |
| 3 | Telemedicine platforms | 20.6% | 18.2% | 22.6% | 20.5% | 3 |
| 4 | Cookware & kitchen goods | 28.0% | 16.1% | 16.1% | 20.1% | 5 |
| 5 | Mental health apps | 22.2% | 17.9% | 16.7% | 18.9% | 4 |
Mattresses tops the list at 44.0%, which makes sense in context. Sleep Foundation, Sleep Advisor, Wirecutter, and Tom's Guide get cited across discovery, shortlist, and variation queries about mattresses with little turnover. Outdoor gear and telemedicine show similar patterns. These are the categories where a content-PR program targeting the small canon publishers compounds across all three buyer intents.
Implications for AI search strategy
Three strategic conclusions follow from the data.
Audit per intent, not per category. A single prompt like "best AI search optimization tool" measures one slice of the citation surface. To get a representative read, run three prompt variants per topic: a discovery prompt, a shortlist prompt, and a variation prompt. The three results will likely disagree, and the disagreement is the signal. If you are visible on discovery but absent on variation, you have a Reddit and marketplace gap. If you are visible on discovery but absent on shortlist, you have an aggregator and comparison-page gap.
Variation queries are a UGC problem. Variation is the intent where "other" citations dominate at 55.8%, driven by Reddit, YouTube, Amazon, and other community or marketplace sources. Vendor-first-party optimization barely moves variation citations. The actual win is being present in the third-party conversations: helpful Reddit threads, YouTube comparison content, marketplace listings with strong reviews. For tech SaaS this often means earning organic Reddit presence (not spam), since AI engines treat Reddit threads as canonical for "cheap alternative to X" and "X for beginners" queries.
Shortlist queries are the aggregator's home turf. Shortlist is the only intent where aggregator citation share (24.7%) slightly outruns vendor share (26.4%). When a buyer asks for a head-to-head, AI engines reach for the comparison page that already exists. The strategic move is to be in the listicles and review aggregators that own those comparisons (G2, Capterra, PCMag, Forbes Advisor, niche publisher hubs in CPG), not to compete with them head-on from your own domain.
The unifying takeaway is that "AI search optimization" is at least three different programs. Discovery rewards listicle inclusion and vendor first-party authority. Shortlist rewards getting into the aggregator comparison universe. Variation rewards community presence and marketplace strength. A team that runs all three optimizes a roughly 3x larger citation surface than a team that runs just one.
Caveats and limits
Five responses per (vertical, intent) cell is a small sample. The headline finding (avg pair-Jaccard at 13.4%, three-intent intersection at 4.3% of union) is robust against this because it averages over 25 verticals, but individual-vertical figures should be read with a wide error band. Some fraction of the observed intent divergence is also driven by prompt-wording noise rather than intent itself; a single discovery prompt per vertical cannot fully separate "what buyers asking discovery-style questions get cited" from "what this exact wording elicits." A larger study (multiple discovery prompts per vertical, multiple shortlist prompts, multiple variation prompts) would tighten the per-vertical estimates, and is on the roadmap for Q3 2026. The taxonomy is hand-curated for the top of the distribution and falls back to heuristics in the long tail; ~10% of citations are unclassified. The data was collected on a single day (2026-05-18) and the engines' underlying training and retrieval indexes drift; this artifact will be refreshed in Q3 2026 with the next benchmark run.
Reproducibility
Every number on this page is derived from the same raw response JSONL used by the Q2 2026 reference benchmark, plus the hand-curated domain taxonomy and a deterministic aggregation script. Identical inputs produce identical outputs.
- Parent reference dataset: AI Search Citation Benchmark, Q2 2026
- Aggregated CSV (
domain × engine × vertical × category): /research/citation-benchmark-2026-q2.csv - Companion analysis (aggregators vs. vendor first-party): When AI Engines Cite the Reviewer vs. the Brand
Want to see how AI engines cite your own site across discovery, shortlist, and variation queries? Run a free Foglift scan with three prompts per topic. Same engines, same grounded production models, applied to your URL.
Frequently Asked Questions
What's a discovery vs shortlist vs variation prompt?
Discovery prompts ask open-ended questions like "Best CRM for a 10-person B2B startup in 2026." Shortlist prompts ask for a ranked list like "Top 10 CRM platforms for small businesses in 2026." Variation prompts add a specific feature or constraint like "Best AI-native CRM in 2026 with built-in pipeline forecasting." The buyer's underlying question changes with intent, and as a result the cited domain set turns over almost completely.
How much does the cited domain set turn over between intents?
The average Jaccard overlap across 25 verticals is 13.4%. That means for the same vertical and same engine, switching from a discovery prompt to a shortlist prompt produces a cited-domain set where only about 13 in 100 domains are shared with the discovery set. Citation strategy that optimizes for one intent does not transfer cleanly to another.
Which intent type is most dominated by listicle content farms?
Shortlist. AI engines reach for listicle content farms (emailvendorselection.com, thedigitalprojectmanager.com, insiderone.com) in 29.6% of shortlist responses versus 28.8% in discovery and just 14.4% in variation. The pattern is consistent across engines: when a buyer asks for a ranked top-N list, AI engines defer to the ranking-formatted SEO sites that have made themselves the canonical answer to those query shapes.
Which intent type is most vendor-first-party-friendly?
Discovery and variation. Vendor first-party share drops 14 points when intent shifts from discovery (66%) to shortlist (52%), because shortlist prompts pull AI engines toward editorial ranking sources rather than vendor pages. If your buyer mostly asks open-ended best-X questions or specific feature questions, your AEO surface is most likely to convert. If they mostly ask top-10-X questions, getting listed in the listicle becomes a much higher-leverage move than optimizing your own pages.
Does the intent gradient hold across all five engines?
Yes, directionally. All five engines show shortlist queries pulling listicle citations harder than discovery does. The exact magnitudes differ. Gemini and Google AI Overview are the most listicle-heavy on shortlist (roughly 45% of responses); ChatGPT is the least (roughly 10%). Claude and Perplexity sit in between. The directional pattern is robust; the per-engine magnitudes are the lever to plan against.
How should this change my content strategy?
Audit the shape of the buyer queries you target. Are they discovery-shaped, shortlist-shaped, or variation-shaped? Then invest accordingly. For discovery-heavy categories, build long-form vendor-first-party content. For shortlist-heavy categories, get listed on the dominant listicle sites and target the #2 vendor share slot through first-party depth. For variation-heavy categories, build deep feature-specific landing pages (think "best AI-native CRM" rather than a generic homepage) since variation queries reward narrow, specific answers.