The sources behind
Hebrew AI answers
For nine months we asked five AI engines 1,645 Hebrew questions, the kind asked before buying a product or service, again and again. That produced 80,528 answers and 489,983 source links. We counted where those links lead.
What was scanned, and what this is not
Five engines, a range of fields, and questions we chose
Every question was asked in each of the five engines, again and again over the months, for small and medium Israeli businesses across a wide range of fields.
The most important limitation, and we say it first: We wrote the questions ourselves, for each business separately, to monitor it. They are not a representative sample of what Israelis search for. This is a dataset of purchase intent questions across a wide range of fields, so every number on this page describes this dataset and not search in Israel as a whole.
Real questions from the dataset
An average of 6.9 words per question. 89.5% open with a question word, 8.8% explicitly ask for a recommendation, 4.2% name a region in Israel and 3.1% ask about price.
In Google you fight for a place on the list.
In AI you fight to be the source the answer is written from.
That is the difference the rest of this page measures
Where the links lead
Seven out of ten referrals go to an ordinary business site
When the 489,983 links are broken down by site type, the big institutions take a much smaller share than people tend to think. Wikipedia, the media and directories together do not reach a tenth.
344,700 links · 16,783 different sites · Anything not in one of the categories below
39,070 links · 285 different sites · gov.il, local authorities and Kol Zchut
33,922 links · 1,347 different sites · org and org.il domains
18,443 links · 16 different sites · Midrag, Dapei Zahav, comparison portals
14,242 links · 11 different sites · Facebook, YouTube, Instagram
13,954 links · 28 different sites · ynet, mako, Globes, Calcalist
11,974 links · 18 different sites · Health funds and hospitals
6,858 links · One site · Links back to Google
3,614 links · 9 different sites · All Wikipedia languages
3,171 links · 172 different sites · ac.il and scientific databases
How the first category is defined: It is the remainder. Any link that is not government, nonprofit, directory, social network, media, public health, Wikipedia, academia or search engine falls into it. A manual check of 400 links, sampled by their actual weight, found fewer than 3%that are not a business, professional or clinic site.
Israeli vs foreign domain
This measures the domain suffix, not the page language, and those are not the same. What it does show: a Hebrew question about a service in Israel mostly returns sources from Israeli domains.
How concentrated it is
- The top ten sites13.8%
- The top fifty sites29.9%
- All the rest, 18,620 sites70.1%
No closed club, but no equality either: 10,856 sites cited three times or fewer together hold 3.4% of the links. The tail is very long and very thin.
Seven out of ten referrals go to the site of someone who does the work.
Not to a giant site, and not to Wikipedia.
70.4% of the 489,983 source links checked
Robustness check
The same finding holds across four ways of calculating
A single number is easy to get by accident, if one big account pulls the average or the collection pipeline skewed it. So we calculated the same breakdown four times, in ways that should break it.
Share of referrals to business sites, within each engine separately
211,283 links · 77.3% of them on an Israeli domain
142,615 links · 76.9% of them on an Israeli domain
39,524 links · 75.0% of them on an Israeli domain
83,566 links · 78.8% of them on an Israeli domain
12,960 links · 56.2% of them on an Israeli domain
The three additional checks
- 71.3% Counting only Perplexity and Gemini, the two engines whose source lists were their own throughout the whole period and were not assembled from search results. 250,807 links.
- 75.2% When repeats are neutralized and each question, engine and site combination is counted once. The repeats actually lower the figure, they do not inflate it.
- 73.4% The median across each scanned business separately, ranging from 41.1% to 90.2%. Median share of Israeli domains per business: 80.2%.
And in each of the nine months the figure stayed within a range of 66% to 74%.
The exception: ChatGPT
Four engines fall within two points of each other. ChatGPT alone points to business sites in 46.6% of the links, and only 56.2% of them have an Israeli suffix. It leans more on institutional and foreign sources. Note that it is also the engine with the fewest links in the dataset, so this is the weakest signal here.
Whoever is not present on the Israeli web
is not present in the Israeli answer.
76.7% of sources are domains with an Israeli suffix
Agreement between engines
Two engines, two source lists
We took every question, source and day combination and checked in how many engines the same source appeared for the same question on the same day. The one day limit matters: without it, an engine that cited a site in January and an engine that cited it in June would count as agreeing.
In how many engines the same source appeared, for the same question, on the same day
163,470 question, source and day combinations
47,281 question, source and day combinations
24,335 question, source and day combinations
The calculation is limited to the three engines that attach sources consistently, so an engine that barely links is not counted as disagreeing.
of sources appeared in only one engine
and only 10.4% appeared in all three. Same question, same day, and three source lists that are mostly strangers to each other.
The practical takeaway: if your site was cited in one engine, do not assume it was cited in the others, and if it was not cited in one, that does not mean it is missing from all of them. This figure is about sources, not name mentions, so it does not directly measure whether the business appears in the answer.
There is no single AI answer. There are five,
and they barely agree on sources.
69.5% of sources appeared in only one engine, for the same question on the same day
What to do with this
Three conclusions you can act on
A business site is a legitimate source
70.4% of the referrals in this dataset go to ordinary sites, not giant ones. A site that clearly explains what the business does, for whom and where, is the kind of source pulled in when an answer is built. That does not guarantee your site will be chosen; it competes for the spot against thousands of others.
Israeli presence is the entry ticket
76.7% of sources are domains with an Israeli suffix. We did not check the language of the pages, but a Hebrew question about a local service almost always returns sources from the Israeli web.
Check several engines, not one
69.5% of sources appear in only one engine on a given day, and ChatGPT behaves differently from the rest. A spot check in one engine tells you about one engine.
Want to know how your site looks to an engine?
Our free check scans a site and tells you whether AI agents find it, read it and understand who is behind it. No sign up and no commitment.
Method, limitations, and what is deliberately left out
- The dataset. The data was collected in the measurement system of the team that develops GEGO, on the same five engines, in real scans for Israeli businesses between January 6 to September 24, 2026. Each question is sent to the engine as is, with no system prompts.
- Selection bias. The questions were written by us to monitor clients, not sampled from search volumes. They lean toward purchase intent and toward the fields in the dataset. No conclusion about what AI engines do in general follows from this.
- What counts as a source. Every address the engine returned in the answer's source list, as it was returned. A domain is counted once per link, including repeat links to the same site in the same answer.
- There is no comparison here of how many sources each engine attaches. We checked, and the numbers came out suspiciously constant for whole weeks, meaning they describe our collection pipeline and not the engine's behavior. The pipeline was replaced on 20.09.2026, and the clean window since then is too small to publish.
- There are no brand mention rates here. Our definition of a mention was tightened in September 2026, so an average over the whole period would mix two definitions. Everything on this page is about sources, not name mentions.
- Privacy. The breakdown is published by category only. There is no ranked list of sites, and no figures on any identifiable business.
- The numbers are correct as of 24.09.2026 and are not updated in real time. The page will be updated when a new, full measurement window has been collected.