The AI Source-Export Gap: A Measurement Finding from Nir Levi’s Israeli GEO Study

Collecting an AI answer does not automatically collect every source the interface displays. During Nir Levi’s Israeli GEO pilot, Perplexity’s on-screen answers grouped several sources behind citation badges, while the Copy control exported representative Markdown links. The study preserved both the copied answer and visible main-panel text. This operational finding matters because the collection method determines what a citation metric can actually describe.

What the pilot recorded

The evidence package preserves 24 complete copied answers and their visible main-panel text. The preselected category brand was mentioned in 24 of those 24 answers; its official domain appeared among the exported links in 22 of 24. Each category contributed four answers. Collection took place on 11 October 2026 in Perplexity Search with the GPT-6.1 Sol model label and fresh incognito conversations.

Source-host overlap is the number of shared hostnames divided by the number of distinct hostnames across the two answer exports. Perplexity’s Copy output is the measured record; grouped on-screen sources may contain additional links.

Name the measured object precisely

The pilot calls its link metric an exported-source measure. It counts the source URLs present in the saved Copy output and compares unique hostnames within that output. This is a reproducible calculation that another reviewer can run against the archived files. It is different from reconstructing every page retrieved by the system or every link hidden inside a grouped badge. A metric becomes more useful when its name tells the reader what evidence was counted.

A brand mention and an owned link can diverge

The first English privileged-access answer mentioned CyberArk in its opening shortlist, while its copied sources included a CyberArk product document hosted at static.carahsoft.com. The export did not include a cyberark.com domain link in that observation. That example makes the distinction concrete: a response can discuss a brand using information distributed through another host. A domain-only report would miss that context unless the analyst retained and inspected the actual page URLs.

Make evidence capture part of the workflow

A useful capture record includes the exact question, language, answer, date, product surface and source export. Preserve a visible interface record when it helps explain grouped citations or other presentation differences. Keep provider-reported source totals separate from exported-link totals. When the interface changes, check the capture method again before comparing a new set of numbers with the old series. This is an operational quality check with direct consequences for the reliability of a GEO report.

Nir Levi’s perspective

First, check which sites are being cited and why. Then identify information missing from the company’s own site and publish useful original content that answers those questions, with real data and examples. The aim is to give AI answers a clearer, better-supported basis for describing the company.

Read the original research

The complete study, paired prompts, evidence downloads and Nir Levi’s commentary are available at https://nirlevi.com/en/research/israeli-geo-hebrew-english/

Put the finding to work

For Israeli businesses reviewing AI visibility with Nir Levi, the lesson is practical: define the evidence first, then calculate the metric. A small, auditable source record is a stronger foundation for a content decision than a larger number whose collection method is unclear.

Explore GEO and AEO consulting in Israel with Nir Levi at nirlevi.com and bring a buyer question to your next review.