Original ResearchBy Sabyasachi Mitter (LinkedIn), Founder and Managing Director at ibs Fulcro · Published August 22, 2026

How to Estimate AEO Prompt Volume: A Calibrated Method

Keyword research tools measure typed queries submitted to search engines. They do not measure prompts submitted to AI assistants. This is the four-stage method ibs Fulcro uses to size Answer Engine Optimisation demand from first-party product data, calibrated against measured search volume, without relying on per-prompt data that no commercial provider currently publishes.

Key Takeaways

Keyword volume is not prompt volume. In a tested Indian automotive category, treating keyword volume as prompt volume overstated addressable AEO demand by approximately 5.6 to 6 times.
The method requires no per-prompt data. It builds a complete prompt universe from first-party product data (specifications, use cases, compliance standards), then calibrates that universe to measured search demand through an explicit, auditable adoption bridge.
AI assistant adoption must be applied per intent, not as a single blended rate. Informational queries have migrated to AI assistants far faster than transactional or local queries; a blended rate misallocates the entire model.

The Problem

Keyword research tools measure typed queries submitted to search engines. They do not measure prompts submitted to AI assistants. Any attempt to size an Answer Engine Optimisation opportunity using keyword volume alone assumes the two are equivalent. They are not, and the error is large: in a tested application across an Indian automotive category, treating keyword volume as prompt volume overstated addressable demand by approximately 5.6 to 6 times.

The method below sizes prompt demand in four stages. It requires no per-prompt volume data, which does not currently exist from any commercial provider. Instead it builds a complete prompt universe from first-party product data, then calibrates that universe to measured search demand through an explicit adoption bridge.

Definitions

Anchor

The entity a prompt is about: a product specification, a use case, a product family, a compliance standard, or a category. Anchors are derived from first-party data, not invented. Anchor count determines the breadth of the prompt universe.

Archetype

A reusable question pattern, such as “best X for Y” or “what specification does Z require”. Archetypes are defined once per layer and applied to every anchor in that layer.

Layer

A set of anchors sharing a question shape, defined by what the buyer knows at the moment of asking. A buyer who knows only their use case asks differently from one who knows the exact product name.

Archetype weight (ω)

A relative score from 1 to 100 expressing how frequently a question shape occurs within its layer. Weights are relative, not absolute volumes. This is deliberate: relative weights survive recalibration, so changing market size or adoption assumptions does not require re-scoring any prompt.

Prompt universe (P)

The total prompt count, calculated as the sum across layers of anchors multiplied by archetypes.

$$P = \sum_l (A_l \times K_l)$$

where A_l is anchor count and K_l is archetype count in layer l.

Stage 1: Construct the Prompt Universe

Define layers by buyer knowledge state. In tested applications this produced six to eight layers. Apply a fixed archetype set to every anchor within a layer. Uniform application matters: writing prompts individually produces uneven coverage that is invisible on inspection, whereas a fixed set applied uniformly makes every gap deliberate and auditable.

Fill variable slots from data rather than assumption. Where a relationship cannot be verified from source data, use fallback phrasing that asks how to determine the answer instead of asserting one. In one build, 12 of 118 vehicle models had no fitment record; those blocks asked how to identify the correct specification rather than stating a specification the data did not support. This discipline is what makes the resulting library defensible.

Stage 2: Establish Measured Search Demand

Begin from total category search volume, then remove non-commercial noise. For any listed manufacturer, a material share of branded search volume is investor and share-price interest rather than purchase intent. In the tested case this was 263,390 of 2,788,490 monthly searches, or 9.4 percent.

$$D = M_{raw} - N$$

Then narrow to the relevant sub-market: $$D_c = D \times s_c$$ where s_c is the category's share of tagged demand.

Stage 3: Bridge Search Demand to Prompt Demand

This is the stage most models omit. Two adjustments are required.

Adjustment one: AI assistant share, applied per intent. Migration to AI assistants is highly uneven by intent type. Informational and research intent has migrated fastest; navigational, transactional, and local intent has barely moved. Applying a single blended rate across all intent destroys this signal and misallocates the entire model.

Recommended AI Assistant Share by Intent Bucket

Intent Bucket	AI Assistant Share (α)
Informational / how-to	0.15 to 0.20
Commercial investigation	0.15 to 0.18
Product type / explanatory	0.12 to 0.16
Transactional / price	0.05 to 0.08
Local / near me	0.03 to 0.05

The gradient is grounded in Seer Interactive's analysis of 49,353 queries (2026), which found AI Overviews trigger on 36% of informational queries, 8% of commercial, and 5% of transactional. Absolute levels are cross-checked against First Page Sage's Q2 2026 estimate placing Google at approximately 80% of all digital queries and ChatGPT at approximately 17%, and against Gartner's February 2024 projection that traditional search engine volume will drop 25% by 2026 as it loses share to AI chatbots and virtual agents. Markets where Google holds near-monopoly share should be weighted below global averages.

Adjustment two: conversational multiplier (μ). One search intent produces more than one prompt, because users refine and ask follow-up questions within a conversation rather than issuing a fresh query. Recommended range 1.3 to 1.5.

This is the weakest parameter in the method, and we want to say so plainly rather than bury it: no published benchmark exists for prompts per intent in any category, and it should be labelled as judgment, not measurement, whenever this method is applied.

$$Q = \mu \times \sum_i (D_c \times w_i \times \alpha_i)$$

where w_i is the share of demand in intent bucket i. Track AI Overview exposure separately rather than adding it to Q: queries that return an AI-generated answer on a search engine represent a genuine citation opportunity but are not prompts, and combining the two double-counts.

Stage 4: Distribute Demand Across the Universe

Allocate Q across layers using explicit shares that sum to one:

$$T_l = Q \times \lambda_l, \quad \sum \lambda_l = 1$$

Critical constraint: layers describe question framing, not intent. Because intent-level adoption was already applied in Stage 3, transactional layers must not be discounted a second time. Layer volumes must never be summed as independent markets: a buyer asking about their use case and the same buyer asking about a specific specification are one person at two moments. Layer totals are allocated shares of a single market.

Within each layer, distribute across anchors using a power law rather than linear decay:

$$f(j) = j^{-\beta} \text{ (rank-ordered anchors), or } f(j) = \left(\frac{c_j}{c_{max}}\right)^{\beta} \text{ (coverage-ranked anchors)}$$$$S_l = \sum_j f(j), \quad V_1 = \frac{T_l}{S_l}, \quad V_j = V_1 \times f(j)$$

Deriving V_1 from the target rather than assuming it guarantees each layer totals exactly to its allocation. Recommended beta: 0.5 to 0.7. The classical Zipf exponent for natural language word frequency approaches 1.0, but empirical query-log studies report consistently flatter exponents because query vocabulary is broader and more evenly distributed than language. Raising beta concentrates volume on head anchors; lowering it flattens the curve. Where no popularity ranking exists, use coverage as the demand proxy: a product available in 18 specifications is assumed to attract more enquiry than one available in 7, mirroring weighted distribution measurement in packaged goods research, where availability proxies realisable demand.

Finally, split each anchor's volume across its prompts by weight:

$$q_p = V_j \times \frac{\omega_p}{\sum \omega}$$

Worked Example: Tyre Category, Single Market, Monthly Figures

Step	Calculation
1	Raw market 2,788,490. Investor noise 263,390. Addressable demand 2,525,100.
2	Passenger car share of tagged demand 35.8%. Category demand 904,188.
3	Intent-weighted AI share: price 58.5% of demand at α 0.08; brand discovery 35.0% at α 0.18; maintenance 2.4% at α 0.20; local 2.2% at α 0.04; specification 1.2% at α 0.20. Sum 107,405. Blended alpha 11.9%.
4	Multiplier 1.4. AEO prompt demand 150,367.
5	Specification layer share 22%. Layer target 33,081.
6	At β 0.55 across 50 anchors, S = 11.30, so V₁ = 2,927. Rank 2 receives 1,999; rank 50 receives 340.
7	Head archetype at ω 100 against Σω 2,817 receives 104 prompts per month.

Seven steps, each inspectable, from a measured market total to a single prompt.

Validation Checks

Four checks should pass before any figure is published

Layer shares sum to 1.0. Summed layer totals equal Q within rounding tolerance. Blended alpha is consistent with published platform share for the market. No layer total is presented as independent addressable demand.

This method in production, applied to a real client: this is the exact methodology ibs Fulcro used to size and validate the prompt universe behind the JK Tyre AEO and GEO audit, where 12,651 calibrated prompts were run across six AI engines. That case study also demonstrates the discipline this framework requires in practice: an initial blended finding overstated JK Tyre's real competitive position, and separating brand-anchored prompts from genuinely generic ones (the same anchor and archetype logic described above) revealed the honest, addressable visibility gap underneath it.

Limitations

Output is modelled demand, not measured per-prompt volume. Relative ranking between prompts is substantially more reliable than any absolute figure. The conversational multiplier and the layer allocation shares are judgment. AI share parameters are third-party industry estimates rather than first-party measurement, and because adoption follows a diffusion curve, any static value will understate demand within a few quarters. Parameters should be revisited quarterly.

Four measurements replace judgment with data, in priority order: assistant-side brand visibility tracking and AI referral traffic replaces alpha; keyword-to-archetype mapping from a matching-terms export replaces omega; conversation session data replaces mu; and search console segmentation replaces lambda.

References

Zipf, G.K. Human Behavior and the Principle of Least Effort (1949). Power-law distribution of ranked demand.
Adamic, L. and Huberman, B. “Zipf's Law and the Internet.” Glottometrics 3 (2002).
Silverstein, C., Henzinger, M., Marais, H., and Moricz, M. “Analysis of a Very Large Web Search Engine Query Log.” SIGIR Forum (1999). Query-log exponent behaviour.
Anderson, C. The Long Tail (2006). Long-tail economics.
Sharp, B. How Brands Grow (2010). Availability as a demand proxy.
Bass, F. “A New Product Growth Model for Consumer Durables.” Management Science (1969). Adoption trajectory.
Seer Interactive. "AIO Impact on Google CTR: 2026 Update." Analysis of 49,353 queries measuring AI Overview trigger rates by intent type.
First Page Sage. "Google vs ChatGPT Market Share: 2026 Report." Q2 2026 estimate of search engine and AI assistant market share.
Gartner, Inc. "Gartner Predicts Search Engine Volume Will Drop 25% by 2026, Due to AI Chatbots and Other Virtual Agents." Press release, February 19, 2024.

If you want our team to build and calibrate an AEO prompt universe for your enterprise, explore our AI Search Visibility (SEO / AEO / GEO) Services.

Let's talk about what this could look like for your brand

Start a Project