Source: ai-research/similarweb-ai-prompt-selection-2026-06-30.md — Similarweb (https://www.similarweb.com/blog/marketing/geo/ai-prompt-selection)
A metric-driven process for deciding which AI-search prompts earn a permanent slot in a GEO tracking campaign — not how to find prompts, but how to triage a candidate list down to the ones that reveal where a brand wins, loses, and can act. Similarweb’s Maayan Zohar Basteker (Jun 17 2026) walks the decision logic inside Similarweb’s AI Search Intelligence, using Uber as the worked example. The core problem it solves: most prompt sets are too big to be useful, so two metrics — citation volatility and citation gap — are used together to keep or cut each prompt.
Key Takeaways
- Prompt selection is a triage step, not discovery. It starts after you already have a candidate list and decides which prompts deserve permanent tracking.
- The default failure mode is a prompt set that’s too big to be useful — pruning is the whole point.
- Two metrics drive every keep/cut decision: citation volatility and citation gap. Read them together, never in isolation.
- Signal 1 — no citations at all means the AI is answering purely from training data; no competitor owns the source slot, so it’s a land-grab opportunity, not a dead end.
- Signal 2 — high volatility with active citations means the AI hasn’t settled on a preferred source yet; the window to become that source is open.
- A citation gap of “None” while still appearing means the brand is in the answer but not driving it — present, not cited.
- Knowing when to leave a prompt out is part of the discipline — not every candidate earns a slot.
Selection vs Discovery
- Discovery (covered separately in Similarweb’s “GEO keyword research” guide) = how to find and research candidate prompts in the first place.
- Selection (this piece) = what happens after you have a list — deciding which candidates earn a permanent slot in the campaign.
- The framing matters because the two need different tooling and judgment: discovery casts wide, selection prunes hard. ^[inferred]
Reading the Signals
Two diagnostic patterns, each tied to a different play. Both are read through the same two metrics — citation volatility (how settled the AI is on its sources) and citation gap (whether the brand is cited vs merely present).
Signal 1 — No citations at all (training-data prompts) → land-grab
- Reading: stable volatility, no citation data, no citation gap flagged → the AI is answering entirely from training data, with no external source competition happening.
- Why it’s an opportunity, not a risk: “no competition” sounds safe but means no entrenched competitor holds the source slot. Publish well-structured, authoritative content first and you become the first cited source by default.
- Playbook:
- Identify which of these prompts connect to a real product moment.
- Create content that answers the question directly and clearly.
- Get there before the citation landscape opens up and someone else claims the slot.
Signal 2 — Volatile prompts with active citations → restructure to be cited
- Worked example (Uber): four prompts at 80% visibility, two at 100%, citations drawn from consumer-services sources.
- Reading volatility: “volatile” citation volatility means the AI hasn’t settled on a preferred source yet — the slot is still contestable.
- Reading the gap: a citation gap of “None” on the 80% prompts means the brand is appearing in responses but not being cited as a source — present in the answer but not driving it.
- Play: the content likely already exists but needs restructuring so AI engines can extract a clear, citable answer:
- Direct answers at the top of the page.
- Structured formatting.
- Explicit answers to the exact question the prompt is asking.
Using the two metrics together
- Citation volatility tells you whether the source slot is still up for grabs (volatile = open; stable = settled).
- Citation gap tells you whether you’re already winning that slot (cited = driving the answer; “None” = present but not driving).
- Read in combination, they sort each prompt into a play — land-grab, restructure, or leave out. ^[inferred]
Try It
- Build a tight prompt set. Start from your discovery list and prune aggressively — a smaller, decision-useful set beats a large one you can’t act on.
- Tag each candidate by the two metrics. Pull citation volatility and citation gap for every prompt before deciding keep/cut.
- Prioritize no-citation prompts that map to a real product moment. These are the land-grab slots — publish authoritative, directly-answering content first to claim the citation by default.
- For volatile-with-citations prompts, restructure the existing page rather than write new content: lead with the direct answer, add structured formatting, answer the exact question the prompt asks.
- Leave prompts out on purpose when neither play applies — selection discipline is keeping the set small.