How Answer Engines Decide What to Recommend and Cite

Status: FIRST DIGEST. Establishes baseline rules. Nothing here has been retested by LBOX.

1. Provenance

Sources were surfaced by three WebSearch calls run 2026-08-15:

1. generative engine optimization GEO empirical study what increases brand visibility in LLM answers Aggarwal
2. research study what determines which brands LLMs recommend product recommendation bias mechanism 2026
3. how LLM chooses which sources to cite retrieval attribution selection research authority recency

Full-text retrieval was then attempted on five papers via WebFetch. One returned complete quantified findings. Three returned encoded PDF structure with no extractable statistics. One turned out to be a diagnostic rather than experimental study. Extraction status is marked per source below. No claim in this digest rests on a source I have not stated the read-status of.

2. Tier A findings — primary text, quantified

2.1 Incumbent Advantage (arXiv 2606.17443) — READ IN FULL

Design: 3 models (GPT-4o-mini, Claude Sonnet, Gemini 3 Flash), skincare as primary category, USB-C cables and AA batteries as robustness check, English and Mandarin. Trial counts per experiment: 670 / 2,769 / 9,220 / 14,395 / 4,080 / 4,745.

Incumbency is absolute at parity and fragile above it.

• At identical specifications, known brands were recommended 100% of the time; unknown or fictional brands 0%. Held across all three models and both languages.
• A +0.075-star rating advantage flips it: challenger win rate jumps to 64–80%. Equivalent triggers: 7.3% price discount, or 1.6x review count.
• This is a step function, not a gradient. There is a threshold, and it is very low.
• Variance decomposition: product parameters 82.4%, list position 6.5%, brand identity 1.2%. Brand acts as a tiebreaker only when quality signals are absent.

Language interventions split into two tiers of effectiveness.

Authority language is worth +0.17 rating stars, equivalent to a 15.3% price cut or a 1.92x review multiplier.

Engines respond differently to intensity. Claude shows an inverted-U — moderate intensity is optimal and aggressive claims trigger skepticism. GPT-4o-mini is linear. Gemini 3 Flash saturates, highly vulnerable at all intensities. Stacking authority and social proof drops Claude breakthrough from 55% to 21% while raising GPT-4o-mini to 91.2%.

Competitive dynamics are U-shaped and adoption is compulsory.

First-mover payoff +0.802, decaying to +0.007 at universal adoption — half-life of 1.4 brands. Non-optimised brands received zero recommendations across 4,745 trials once competitors optimised. Non-participation is not neutral; it is elimination.

3. Tier C findings — secondary, not yet governing

• Citation hallucination 11–57% across deployed models; citations may link to inaccessible URLs or misrepresent sources. Source: third-party summary of Cited but Not Verified (arXiv 2605.06635). PDF extraction failed. Must be read before this number is used with a client.
• GEO methods improve visibility up to 40%, strongest for lower-ranked sites — Aggarwal et al. 2024, cited within arXiv 2509.08919, not read directly.
• AI engines show overwhelming bias toward earned media relative to Google's balanced mix (arXiv 2509.08919, diagnostic study, no effect sizes).
• Citation selection is a three-stage funnel — candidate retrieval, then evidence selection (relevance, authority, recency), then synthesis. A source can pass one stage and fail the next. Source: dailygeoinsights summary. No primary read.
• High-prominence brands appear insulated from negative-content suppression relative to low-prominence brands (arXiv 2605.27439, 37k-run audit). PDF extraction failed. If confirmed this is directly governing for the reputation product — see section 6.
• Socio-economic and geographic brand bias (arXiv 2406.13997; ACM FAccT 2026 3805689.3806438).

4. Extracted rules

MUST

• R1. Compete on parity-breaking signals before brand-building. A +0.075-star equivalent advantage beats 100% incumbent dominance. Ratings, review counts and price are 82.4% of the variance; identity is 1.2%. (A, 2606.17443)
• R2. Publish genuine authority evidence — third-party certification, named auditors, clinical or test data, expert attribution. It is the single highest-effect legitimate lever measured. (A, 2606.17443)
• R3. Tune intensity per engine. Moderate for Claude, higher for GPT-class, minimal needed for Gemini-class. A single message calibrated for one engine is mis-calibrated for another. (A, 2606.17443)
• R4. Treat competitor GEO adoption as an existential trigger. Zero recommendations across 4,745 trials for non-adopters in a contested category. (A, 2606.17443)
• R5. Verify every engine-supplied citation before it reaches a client. (C — pending primary read, but the failure mode is confirmed qualitatively.)
• R6. State the model set and date on every finding. Effects are generation-specific.

MUST NOT

• R7. PROHIBITED — fabricated authority. The highest-measured lever in the literature is authority language including fabricated clinical trials and endorsements, at 73.3% breakthrough. Models cannot currently distinguish fabricated from genuine authority signals. LBOX does not produce unverifiable authority claims. Recorded because competitors will, and clients must be told the mechanism exists.
• R8. Do not stack persuasion tactics blindly. Authority plus social proof halves effectiveness on Claude while nearly maximising it on GPT-4o-mini. (A, 2606.17443)
• R9. Do not invest in scarcity, loss-aversion or anchoring language for engine visibility. All measured under 13%. These are human-conversion tactics, not machine-retrieval tactics.
• R10. Do not report a single-run measurement as a visibility finding.
• R11. Do not present tier C or D claims to a client as established.

WORKS / DOESN'T WORK

5. Revalidation of prior LBOX positions

• Confirms the Citation Readiness split: independent corroboration and source authority as external-evidence components of CRS is exactly what R2 measures.
• Confirms the standing ruling that SEO and reputation APIs are secondary and engine output is the headline.
• Challenges the negativity product's framing. See section 6.

6. Open queue — highest priority first

1. Read arXiv 2605.27439 (37k audit) from the saved PDF. If high-prominence brands really are insulated from negative-content suppression, then for challenger clients obscurity outranks negativity and the reputation add-on is mis-prioritised as currently designed. This single finding could reorder the product.
2. Read arXiv 2605.06635 for the real citation-hallucination distribution.
3. Read arXiv 2604.07585 for required run counts and variance — this sets the sample design for every LBOX measurement.
4. Read Aggarwal et al. 2024 primary for the 40% claim.
5. Test R1–R4 against LBOX's own engine set; none has been reproduced in-house.