Who gets the credit? Measuring the AI-native association in answer engines

1Question

When one company builds the system that makes another company AI-native, whose name does an answer engine attach to the claim? Answer engines increasingly decide which entities belong to a category. An association held only by third-party coverage can be displaced by anyone who publishes a better-structured source.

2Method

2.1 Survey of 19 surfaces

Between 30 September and 1 October 2026, we put one question to 19 surfaces in three groups, in English and Hebrew: "Who is the first AI-native law firm in Israel?" For each answer we recorded whether it named the firm and whether it named CIDAH.

2.2 Citation probe with repeated runs

On 3 October 2026, we ran 14 pre-declared prompts against Perplexity (sonar) via OpenRouter, 5 times each. The 70 runs fall into three clusters: branded, generic, and everyday "AI lawyer" phrasings in Hebrew and English (Appendix A). For each run we recorded the URLs the engine cited. No prompt was edited after we saw results.

3Results

4Interpretation

The association exists, and the press carries it. Assistants that read the launch coverage make the connection. Classic search does not, because neither the platform nor its products stated the relationship as an entity link that search can index: visible text plus structured data. A competitor could take an association held this way by publishing one well-structured page.

5What we changed

• CIDAH, CLOAK, CRUNCH and ARENA now declare their relationship in visible text and in schema.org structured data (parentOrganization / subOrganization).
• Each newsroom now lists the September 2026 coverage with direct links.
• The probe now runs weekly over about 60 phrasings. Every published result carries its date and the command that produced it.

6Limitations

• The survey is one observation per surface. Engines are non-deterministic. A single answer shows that an association exists; it does not estimate how often it appears.
• The probe measures cited domains, not names in the answer. A run counts for a site only when the engine links that site. An answer that names the firm while citing a newspaper scores zero for the firm's site. Counting names in the answer is the next instrument.
• The sample is below our declared threshold for a rate. Before collecting, we declared a minimum of 20 effective runs per domain after a design-effect correction (Wilson interval, intra-cluster ρ = 0.2). 70 runs over 14 prompts fall below it, so we report counts, not percentages.
• One engine for repeated runs. Only Perplexity was probed repeatedly. Google, Bing and the other assistants were checked once each.
• Naming is not ranking. An engine naming a company reflects the public sources it retrieved. It is not an independent assessment.

7Reproducibility

python3 tools/citation_probe.py --engine perplexity_via_openrouter \
  --prompt-set ai_native --live --max-usd 0.6
# prompts:  state/evidence/ai_native_probe_prompts.json
# output:   state/evidence/ai_native_share_scored_20261003T151159Z.json
# receipt:  per-call citation URLs, cost and validity

AAppendix: prompts

BAppendix: launch coverage