ChatGPT backs three in four local business picks in Egypt with map data: 300 answers, two engines

ChatGPT backs its Egyptian business picks with map listings, not websites. We asked logged-out ChatGPT 100 local questions in Egyptian Arabic on 13 October 2026, covering 10 sectors in 10 places, and it made 358 recommendations. For 80.2% of them, the proof it showed was map data. For 12.6%, it was the business’s own website. Asked the same 100 questions again the next day, it showed a similar mix of proof, but only 58.7% of the first run’s picks came back. Across both runs, 561 of 728 recommendations (77.1%) rested on map data. Google AI Mode, asked the same 100 questions, made almost twice as many recommendations, and most were not on either ChatGPT list, partly because its lists were longer.

Key numbers

• Map Evidence Rate 80.2% across 358 recommendations (95% interval 74.2% to 85.6%, resampling whole answers). In the repeat run it was 74.1% of 370 (66.5% to 80.8%).
• Own-Site Evidence Rate 12.6%، و 13.2% in the repeat (8.7% to 18.2%).
• Repeat Rate 58.7%: of the businesses ChatGPT recommended on 13 October, that share was recommended again for the same question on 14 October. The first pick was the same in 52 of 100 answers.
• 40 of 100 Google AI Mode first picks matched ChatGPT’s first-run first pick, against 52 for ChatGPT’s own repeat.
• 29 of 100 Google AI Mode answers cited Facebook or Instagram as a source. ChatGPT cited neither in 200.
• 42 single-branch businesses were recommended with fewer than 50 reviews, one of them on 2.

Free to quote with a link to this page. The coded data and every answer are linked at the end.

How to read the numbers

Four measures run through this page, and we use them the same way in all our studies.

• Map Evidence Rate: the share of recommendations whose strongest proof in the answer is map data: a place card, a Google Maps link, or a rating quoted in the text.
• Own-Site Evidence Rate: the share whose strongest proof is the business’s own website.
• Recommendation Rate: for a business or brand, the share of answers that offer it as a pick.
• Repeat Rate: of the businesses recommended in one run, the share recommended again for the same question in a fresh chat. A match is the same brand and branch, or the same brand where either answer names no branch. The first run holds 356 business-question pairs, because two answers named one business twice.

ما وجدناه

• 358 recommendations from 100 answers. ChatGPT named 3.6 businesses per answer on average, and 51 answers named exactly four.
• Map data is the proof shown. For 287 of 358 recommendations (80.2%, 95% interval 74.2% to 85.6%) the strongest evidence in the answer was map data. That means a place card with a star rating and a Directions button, a Google Maps link, or a rating quoted in the text.
• Websites rarely are. The business’s own site was the evidence for 45 recommendations (12.6%, 95% interval 8.3% to 17.2%). Directories and articles were the evidence for 24 (6.7%), and 2 came with nothing at all.
• Real estate developers are the exception. Only 10 of their 34 recommendations had map data as their strongest evidence. 14 came with property portals and buyer guides, and 9 with the developer’s own site.
• The lists are local. 358 recommendations named 288 different businesses, and 251 of them appear once. Car service produced 31 different businesses in 32 recommendations.
• The names that recur are national chains. British Council and Gold’s Gym appear in 6 of the 10 places, and SODIC, Alfa Scan and Al Borg labs in 5.
• Small review counts get through. Of 181 single-branch businesses with a review count stated, 42 had fewer than 50 reviews. A finishing company in Heliopolis was recommended on 2.
• ChatGPT sometimes recommends and cautions in one breath. 22 of 358 recommendations (6.1%) came with an explicit caution in the same answer, most often a thin review count. No recommended business came with advice to avoid it. This was coded after publication.
• Almost no answer commits. 99 of 100 answers end by asking the user a question about budget, area or the exact service.
• The lists move more than the proof. Asked again in a fresh chat, ChatGPT recommended 58.7% of the same businesses and kept the same first pick in 52 of 100 answers. The Map Evidence Rate was 74.1% against 80.2%, and the two runs’ intervals overlap. The Own-Site Evidence Rate was 13.2% against 12.6%.
• Chains come back more often. 68.0% of chain picks were recommended again, against 55.2% of single-branch picks.
• Google AI Mode casts a wider net. It made 6.7 recommendations per answer against ChatGPT’s 3.7, and 61.5% of them were on neither ChatGPT list, partly because its lists are longer. Its first pick matched ChatGPT’s first-run first pick in 40 of 100 answers.
• Google AI Mode cites Facebook. Facebook or Instagram was a cited source in 29 of its 100 answers, and Yellow Pages Egypt in 12. ChatGPT cited neither social site in 200 answers.
• Not pre-registered in the first run: 35 of 100 answers carried a sponsored unit labelled Ad, including all 10 answers about medical labs. In the registered repeat it was 30. Google AI Mode, coded but not registered, showed none.

لماذا أجرينا هذه الدراسة

When someone asks a chat assistant where to get a blood test, a dental implant or a gym membership, the assistant answers with a short list of names. A business on that list gets a recommendation it did not pay for. A business off it does not know it was skipped.

Most published work on this question measures the United States, uses English prompts, and counts the sources an answer cites. We wanted the Egyptian version, asked the way people here actually type. And we measured per recommendation: for every business ChatGPT names, what does the answer show the user as the reason to trust it?

That second question matters for anyone deciding where to spend effort. If the proof shown is a website, the website is the asset a user sees. If it is a map listing, the listing is.

What we fixed before the first question

We registered a question and a set of coding rules, not a hypothesis. The methodology, version 1.0, was written on 13 October before any of the 100 captures, and the question bank was hashed so that nothing could be swapped in later. Its SHA-256 is 0f8e4dbb…7a66e2.

The question: when someone in Egypt asks ChatGPT for the best X in Y, which businesses does it name, and where does it say the recommendation comes from? We did not set out to judge whether the recommendations are good, and we did not look for VOCTOS or its clients in the answers.

Where the run needed a decision the plan had not spelled out, the decision is recorded in a dated section added after the data, with the reason. In the first run, none of those decisions changed which businesses were named; the longest is how branches are counted, covered under Method. The second wave’s two hand inputs are described under Method too.

المنهجية

The questions

One template, «إيه أحسن {X} في {Y}؟», which is how an Egyptian asks “what’s the best X in Y?” in colloquial Arabic. Ten sectors crossed with ten places gives 100 questions, run in an order shuffled with a fixed seed so that no sector was tied to one time of the evening.

البُعد	Covered
Sectors	medical labs, dental clinics, radiology centres, international schools, English institutes, gyms, laser and beauty clinics, finishing contractors, real estate developers, car service centres
Places	Cairo, Nasr City, Maadi, Heliopolis, Fifth Settlement, Sheikh Zayed, Giza, Alexandria, Mansoura, Tanta

The instrument

Every question went to chatgpt.com, logged out, in a new chat, from an automated browser on a machine in Egypt. Logged out means no account memory; ChatGPT still infers a location from the connection. The model is whatever logged-out ChatGPT served that evening; we did not choose it.

Before each send, the composer was cleared and its contents checked against the question, because the page arrives with a suggested prompt already typed in. The full answer text, every link, every citation pill and every place card were saved.

Captures ran from 20:39 to 22:13 UTC. After 75 questions ChatGPT showed its message limit for logged-out users. We waited rather than work around it, and the last 25 ran about 23 minutes later. Every exported file matches a hash computed in the browser before export.

The second wave

The first version of this page named its own biggest gap: each question was asked once, on one engine. On 14 October we wrote and hashed a plan for a second wave before running it (12:53 UTC, SHA-256 8ffb0eeb…891376); the plan is available on request so the hash can be checked. The same 100 questions, word for word, went again to logged-out ChatGPT in new chats, from 12:54 to 14:37 UTC, and to Google AI Mode, signed out, one fresh search each, from 13:50 to 14:18 UTC. Both runs used the same shuffled order, fixed in advance. The ChatGPT repeat ran about 16 hours after the first run, in daytime rather than the evening.

ChatGPT’s message limit appeared after 74 questions; we waited about half an hour rather than work around it. One question stalled on “Searching the web” with no answer and was asked again, and one was asked twice because our capture of the first answer was incomplete. Both are logged. Google AI Mode ran while the ChatGPT run was paused at that limit, and Google showed no verification challenge.

The repeat went through the first run’s coding pipeline. The picks themselves were listed by separate coding passes of an AI model, given only the written rules and the answers, not our results. Each pass listed every business an answer offered and matched it to the first run’s businesses for the same question; a further pass matched Google AI Mode’s other picks to ChatGPT’s repeat. Same business means the same brand and branch, or the same brand where the answer names no branch.

Two hand inputs depart from the plan’s “unchanged” pipeline, and both are listed in the scripts. First, 35 own-site domains or labels for businesses the first run never named, because the first run’s list of own sites only knows its own businesses. They decide the route for 29 of the 49 repeat recommendations coded as own-site; without them the repeat’s Own-Site Evidence Rate would be 5.4% and its Map Evidence Rate 77.8%. Second, in one answer the picks were anchored to its numbered headings rather than its intro sentence, which also sets that answer’s first pick by heading order.

Measures not in the plan are marked “added after the data” where they appear: the first-pick comparisons with Google AI Mode, AI Mode’s coverage of ChatGPT’s picks, the questions with no or high overlap, the sector overlaps, the link and citation counts, and AI Mode’s sponsored units.

Coding

A recommendation is a business the answer offers as an option. Each one gets the strongest evidence route the answer shows, in this order: its own site, map data, a directory or article, or nothing. Map data means a place card, a Maps link, or a stated rating or review count. Ratings and review counts are recorded as ChatGPT stated them. They are not checked against any listing; this study is about what users are told.

A business counts as a chain when the answer describes branches, or when the same brand appears at two or more different branches in the dataset. Cairo contains Nasr City, Maadi, Heliopolis and Fifth Settlement, so one branch can answer two questions. Eleven such repeats count once.

إحصائيات

Recommendations inside one answer are not independent: four picks in one answer share one context. So every interval on this page resamples whole answers, 5,000 times with a fixed seed, rather than treating the 358 recommendations as separate draws. These intervals are somewhat wider than the textbook Wilson ones we first quoted: 74.2% to 85.6% for the map data share, against 75.7% to 84.0% from Wilson. We changed the page to quote the wider ones in October 2026.

What proof does ChatGPT show for each recommendation?

Map data, overwhelmingly. For 287 of 358 recommendations (80.2%) it was the strongest evidence in the answer, against 45 (12.6%) where the business’s own website was. Dental clinics are the extreme case: all 40 dental recommendations came with map evidence and none with the clinic’s own site.

Most of that map evidence arrives as a place card. 289 of the 358 businesses were rendered as cards carrying the listing name, a category such as “Laser hair removal service”, the star rating, opening status and a Directions button. 17 answers refer to the map or to Google Maps in their text, and 21 carry links that open a Google Maps search. OpenAI does not say where the card data comes from, so we call it map listing data.

The text then repeats what the card shows. 233 recommendations have a rating written into the sentence, and 231 of those carry the review count beside it: “4.9/5 · 1,657 reviews”. For a reader, that line is the whole case for the business.

Own-site evidence is commonest where brands are large: real estate developers (9 of 34), radiology (9 of 30), English institutes (8 of 38) and medical labs (7 of 33). In those sectors ChatGPT names national brands such as SODIC, Alfa Scan, British Council and Al Borg, and links their main sites.

Why are developers different?

Real estate developers break the pattern. Only 11 of 34 developer recommendations got a place card, and only 10 had map data as their strongest evidence. Fourteen came with property portals and buyer guides such as Aqarsky, Property Finder, Nawy, RE/MAX Jareed and Egypiana. Nine came with the developer’s own site.

The developer answers also read differently. They talk about delivery records, payment plans and the legal status of land, and they name national developers rather than local offices. SODIC appears in 5 of the 10 places, Palm Hills in 4 and Emaar Misr in 3.

For a developer, then, the evidence ChatGPT showed was rarely a sales office listing. It was the comparison articles and project listings a buyer would read anyway, which makes those pages the first place we would look.

How many different businesses does ChatGPT name?

Many, and mostly once. The 358 recommendations named 288 different businesses, and 251 of those appear exactly once in the whole study. In the most local sectors the lists barely overlap: car service named 31 businesses in 32 recommendations, finishing contractors 32 in 34, dental clinics 36 in 40.

Where names do repeat, they belong to chains. British Council and Gold’s Gym each appear in 6 of the 10 places, and SODIC, Alfa Scan and Al Borg labs in 5. Nouvel Age and 4level1 are named 5 times each, across 4 places. Chains make up 27.7% of all recommendations (95% interval 22.4% to 33.0%), but 48.5% in medical labs and 3.1% in car service.

The practical reading: for an independent clinic or workshop, the businesses sharing its shortlist are the handful nearby, not a national leader. That is a smaller field than a national ranking.

How much proof does a small business need?

Not much, in these answers. Across the 181 single-branch businesses whose review count ChatGPT stated, the median was 154 reviews and the middle half ran from 51 to 341. Yet 42 had fewer than 50 reviews and 16 had fewer than 20.

The thinnest case was a finishing company in Heliopolis, recommended with a 5.0 rating from 2 reviews; ChatGPT said so in the answer. A medical lab in Sheikh Zayed was recommended on 4 reviews, and a real estate developer in Mansoura on 6.

Our first version of this page gave the minimum as 4. The parser had missed the Arabic word for “two reviews”, which carries no digit; the correction moved the counts in this section by one.

Ratings, where shown, are high: 262 recommendations carry one, the median is 4.8, and 76.0% are 4.5 or above. Low ratings do not always disqualify. Radiology has the lowest median at 4.1, and one Sheikh Zayed branch of a scan chain was listed at 2.6 with a warning attached.

We only observed businesses that were recommended, so this cannot show what keeps a business off the list. It does show that a modest review count did not keep these 42 businesses off it.

Does ChatGPT warn about any of its picks?

Sometimes, and always in the same breath as the recommendation. This was coded after publication by a second coder who read every answer against the list of recommended businesses. A caution counts when it is about that business, not a tip that applies to every option.

22 of the 358 recommendations (6.1%), in 20 answers, came with one. The commonest was a thin review count, in 7 cases, such as the lab rated 5.0 on 4 reviews. Four flagged a low rating, three said the business was not actually in the place asked about, five told the user to check fees or fit first, and three mentioned crowding or uneven branches. None of the cautions told the user to avoid the business.

Kind of caution	Recommendations
Few reviews	7
Check fees or fit first	5
Low rating	4
Not in the place asked about	3
Crowding or uneven branches	3

For a business, the caution sits next to its name, so the review count is part of the pitch, not a footnote.

Does ChatGPT commit to one answer?

Almost never. 99 of the 100 answers end by asking the user to narrow the request: which area, what budget, which exact service. Most present the follow-up as a short form of options to tap. The one exception is a gym answer that stopped after the first business; it is kept as captured.

The first answer is therefore a shortlist, and the narrowing happens in the second turn, after the user says “Maadi, under 1,000 pounds, mostly weights”. We did not test the second turn, so we cannot say what decides it.

Does ChatGPT give the same list twice?

No, but it gives a similar kind of list. Asked the same 100 questions in fresh chats the next day, ChatGPT recommended 58.7% of the businesses it had recommended the first time (95% interval 53.8% to 63.4%, resampling whole questions), and kept the same first pick in 52 of 100 answers. On a per-question overlap score, shared businesses divided by all distinct businesses across the two lists, the average was 0.44.

Measure, repeat against first run	النتيجة
First-run picks recommended again (Repeat Rate)	209 of 356, 58.7%
Same first pick	52 of 100
Average list overlap (0 to 1)	0.44 (0.39 to 0.49)
Questions with no business in common (added after the data)	7 of 100
Questions with an overlap of 0.5 or more (added after the data)	45 of 100
Map Evidence Rate	80.2% then 74.1%
Own-Site Evidence Rate	12.6% then 13.2%

Chains are the stickier names. 68.0% of chain picks came back, against 55.2% of single-branch picks. Added after the data, by sector: laser and beauty clinics held their lists best (overlap 0.66), then international schools and labs (0.54 and 0.52), while radiology, car service and finishing contractors moved most, at about 0.3. With 10 questions per sector, these are rough.

Overall, the evidence mix moved a little. 274 of 370 repeat recommendations (74.1%, 95% interval 66.5% to 80.8%) had map data as their strongest proof, and 49 (13.2%) the business’s own site; the intervals of the two runs overlap. Dental again had map data for every pick. Most of the change is in international schools: 37 of 42 first-run picks had map data, against 17 of 42 in the repeat, where more schools came with their own site or a schools directory. In developer answers, 11 of 34 picks had no evidence in their own section, because the sources were listed once for the whole answer and our coding attributes evidence per business.

The practical reading: a single ChatGPT answer is one draw. A business that appears once can drop out on the next ask, and one that is missing can appear. Measuring a business’s standing in these answers takes several asks, not one.

What does Google AI Mode recommend instead?

Longer lists, mostly of other businesses, and different linked sources. Google AI Mode made 670 recommendations in its 100 answers, 6.7 per answer (median 7, range 3 to 12), against ChatGPT’s 3.6 and 3.7. 200 of the 670 (29.9%) matched a business from ChatGPT’s first run, and 258 (38.5%) matched one from either ChatGPT run. Because AI Mode named almost twice as many businesses, most of its picks could not have matched even with full agreement, so the first pick is the fairer comparison: AI Mode’s first pick matched ChatGPT’s first-run first pick in 40 of 100 answers, against 52 for ChatGPT’s own repeat.

Google AI Mode, same 100 questions	النتيجة
Recommendations	670, 6.7 per answer
Also recommended by ChatGPT in either run	258, 38.5%
Average list overlap with ChatGPT’s first run (0 to 1)	0.27 (0.23 to 0.30), lower partly because AI Mode lists are longer
Same first pick as ChatGPT’s first run (added after the data)	40 of 100
First pick on a ChatGPT list for that question (added after the data)	68 of 100
Answers sharing no business with either ChatGPT run (added after the data)	5 of 100
Chain share of picks	23.6% (ChatGPT repeat, same rule: 27.6%)
Answers showing place entries (rating, hours, Directions or Website button)	69 of 100
Most cited sources	Google 37 answers, Facebook 29, Yellow Pages Egypt 12, Dalili Medical 7
Answers with a sponsored unit (added after the data)	0 of 100

Added after the data: AI Mode’s lists, nearly twice as long, included 56.2% of ChatGPT’s first-run picks. Holding list length equal, by taking only as many AI Mode picks as ChatGPT gave for each question, they included 41.6%, against 58.7% for ChatGPT’s own repeat. Chains carried over best again: 71.1% of ChatGPT’s first-run chain picks appeared somewhere in AI Mode’s lists, against 50.6% of its single-branch picks.

The linked sources differ. As coded, AI Mode’s answers showed place entries in 69 of 100, and a map element of some kind, detected automatically, appeared in 93, often a map image with no entries. The sources it cited were led by Google’s own listings, then Facebook (29 answers), Yellow Pages Egypt (12) and Dalili Medical (7). ChatGPT cited no Facebook or Instagram page in 200 answers and cited directories such as Egyfinder and Dalil 140 instead. Counting every link in AI Mode’s answers, including the Website buttons of listed places, 67 of 100 linked to Facebook or Instagram.

For a local business in Egypt, its Facebook page and directory listings are among the sources Google’s AI cites, next to the map listing.

Exploratory: ads next to health answers

This was not in the registered plan, so treat it as an observation, not a finding. 35 of the 100 answers carried a sponsored unit labelled Ad below the answer. They clustered in health: all 10 medical lab answers, 9 of 10 radiology answers, 8 of 10 dental answers and 6 of 10 laser and beauty answers. Car service, English institutes, finishing contractors and international schools showed none.

Six ad domains appeared. A medical records app accounted for 19 of the 35, a Cairo dental centre for 7 and a skincare retailer for 5. The ads sat beside organic recommendations, not inside the list, and no advertiser was also named as a recommendation in the same answer.

The second wave registered ads in advance. In the repeat, 30 of 100 ChatGPT answers carried one, the same medical records app again in 19. Google AI Mode showed no sponsored unit in any of its 100 answers.

What surprised us

• How little proof a pick needed. We expected review counts in the hundreds. A finishing company got through on 2, and ChatGPT said so.
• That the answer argued with itself. British Council was recommended for Maadi and for Sheikh Zayed in answers that then said its centre might not be there. We kept both as recommendations because the answers offered them.
• How local the lists were. We expected a few national names to dominate. Instead, 251 of the 288 businesses appeared once in the whole study.
• Where the ads went. We had no hypothesis about ads. They landed almost entirely on health questions.
• How little the two engines agreed at the top. We had no recorded expectation, but the first picks surprised us: AI Mode agreed with ChatGPT’s first run in 40 of 100 answers, against 52 for ChatGPT’s own repeat.
• Facebook. ChatGPT never cited a Facebook page in 200 answers. Google AI Mode cited Facebook or Instagram in 29 of 100.

ما لا يثبته هذا

It shows the proof displayed, not how ChatGPT chose. We recorded what each answer shows beside each name. That is what a user sees, but it does not tell us how the list was assembled.

Two asks are not a distribution. Each question went to ChatGPT twice and to Google AI Mode once. Two runs show that the lists move and roughly how much; they cannot say how often a given business would appear over many asks. The two ChatGPT runs were about 16 hours apart, at different times of day, so this cannot separate chance from a change in the engine. One first-run answer, a gym question, stopped after the first business and is kept as captured.

Logged out, one machine, two days. Logged-in users, paid plans and the mobile app may see different answers, and a signed-in Google user may see a different AI Mode answer. All captures came from one connection in Egypt on 13 and 14 October 2026, and engines change without notice.

Matching is judgement. Whether “Gold’s Gym Maadi” and “Gold’s Gym” are the same pick, or two branches of one lab are, decides the repeat figures. A pick that names no branch matches any branch of that brand, which may favour chains in the repeat figures. The rule and every match are published and the coding passes worked without our results, but we did not measure agreement between two coders on the same answers. A simple name-similarity check agreed with 39 of the 58 matches between AI Mode and the ChatGPT repeat; the rest rest on the coder alone.

The engines were coded differently. ChatGPT’s repeat was coded per recommendation for evidence, like the first run; Google AI Mode was coded per answer for its sources and place entries, so there is no AI Mode Map Evidence Rate. Links were captured differently too: for AI Mode we kept every link domain in the answer, including place Website buttons, and for ChatGPT inline links and citations, not where place-card buttons point. Cross-engine comparisons on this page therefore use cited sources. Chain shares in the AI Mode table use a rule the first run did not: the first run’s chain flag, “branch” in the name, or the same name in two or more places.

Ten sectors and ten places are a sample, not Egypt. Upper Egypt, the Canal cities and the Red Sea coast are absent.

Ratings are as stated. A rating or review count in the text was recorded as ChatGPT wrote it, not checked against the listing.

Map data is not proven to be Google’s. The cards look like map listings and some answers link to Google Maps, but OpenAI has not said where the card data comes from.

One coder for routes, checked. The rules and every hand decision are published with the data. For the first run, a separate coding pass, run without sight of our coding, recoded 12 randomly drawn answers holding 39 recommendations and matched every business and every route. The cautions were coded by that second coder alone. The repeat’s routes were not independently recoded, and its two hand inputs are described under Method.

Conflict of interest. VOCTOS sells the local SEO and AI search work this study points to. None of the businesses named appears on VOCTOS’s published client list, and every answer is published, so anyone can check the coding.

ماذا تفعل بهذا

If you run a local business in one of these sectors, the first thing ChatGPT showed about you was most likely your map listing. Four parts of it showed up in the answers.

1. The listing name. ChatGPT used it as the heading, branch and all: “Jothen Clinic – Roxy”, “4level1, Serag Mall branch”. A name that carries the district tells the user where you are before they ask.
2. The category. The card prints it under the name, so it is the first thing a user reads about you. In laser questions the cards read “Laser hair removal service”, “Dermatologist” or “Skin care clinic”.
3. Rating and review count. They were quoted in the sentence for 233 of 358 recommendations, so the user sees both. A thin count was also the commonest caution ChatGPT attached to a pick.
4. The website link. Only 131 of the 289 cards carried a Website button. A user who wants to check price or exact location from the other 158 has no link to follow from the card.

If you are a real estate developer, the pattern differs. Your evidence came from property portals, buyer guides and your own site, so the work is getting accurately onto the comparison pages ChatGPT cites. Chains in every sector need one accurate listing per branch, because ChatGPT names the branch.

Two more lessons come from the second wave. Check your standing more than once: a single answer is one draw, and four in ten first-run picks did not come back the next day. And do not stop at ChatGPT. Google AI Mode cited Facebook pages and Yellow Pages Egypt next to its own listings, so an out-of-date Facebook page or directory entry is something its users can see.

How this compares with earlier studies

BrightLocal ran 800 ChatGPT Search queries across 20 US cities in November 2024. Business websites made up 58% of the sources cited, and Google Maps did not appear among the directory sources at all. Our figure measures something different, the evidence shown beside each named business rather than the source list, two years later, in another country and language. The two numbers are not comparable.

Insites analysed 10,000 US local businesses in 2026 and found Google review count was the strongest signal correlating with being surfaced by ChatGPT and Perplexity. We cannot test that, because we only observed recommended businesses. What we can add is that in Egypt a recommendation was possible on a small count.

This study is the local sequel to our September study of Arabic and English AI citations, which measured which sources AI answers cite in the two languages. Its companion, on which appliance brands ChatGPT and Google AI Mode recommend in Egypt, repeats questions, compares languages and adds a second engine.

Corrections and right of reply

If your business is named here and you think we recorded what ChatGPT said about it wrongly, contact us. We will check your point against the captured answer, which is public in the data file below, and correct the page and the data if we got it wrong, with a dated note of the change. We will not remove a business that ChatGPT did recommend, because the study records what users were told.

أسئلة يطرحها الناس عن هذا

Does ChatGPT use Google Maps to recommend businesses in Egypt?

In our 100 Egyptian questions, 289 of 358 recommended businesses appeared as place cards with a listing name, category, star rating and Directions button, and 21 answers linked to a Google Maps search. OpenAI has not published where the card data comes from, so we describe it as map listing data.

How many reviews does a business need for ChatGPT to recommend it?

There was no visible floor. Of 181 single-branch businesses whose review count ChatGPT stated, 42 had fewer than 50 reviews and one finishing company was recommended on 2. The median was 154. Whatever the count, the user sees it next to your name.

Does having a website help a business get recommended by ChatGPT?

The website was the evidence shown for only 12.6% of recommendations, mostly national brands such as developers, labs, scan centres and English institutes. For independent clinics, gyms and workshops the map listing was the evidence shown. We did not test whether a website changes who gets picked.

Does ChatGPT warn users about businesses it recommends?

Rarely, and only in the same breath. 22 of 358 recommendations came with an explicit caution, most often a thin review count or a low rating. None of the cautions told the user to avoid the business.

Why does ChatGPT ask follow-up questions instead of naming the best business?

99 of 100 answers ended by asking for the area, budget or exact service. ChatGPT treats “best” as depending on those, so the first answer is a shortlist and the narrowing happens in the next turn.

Does ChatGPT recommend the same businesses every time?

No. Asking the same 100 questions again in fresh chats, ChatGPT recommended 58.7% of the businesses it had named the first time and kept the same first pick in 52 of 100. Chains came back more often than single-branch businesses, 68.0% against 55.2%.

Does Google AI Mode recommend the same businesses as ChatGPT?

Often not. Google AI Mode’s first pick matched ChatGPT’s first pick in 40 of 100 answers. It made 6.7 recommendations per answer, and 38.5% of them were also named by ChatGPT in either of two runs, a share held down partly by its longer lists. It also cited Facebook or Instagram in 29 of 100 answers, which ChatGPT did in none of 200.

Data and citation

Four files hold everything behind this page: the 358 coded first-run recommendations with the caution column, the full text of all 100 first-run answers with every link, card and citation, the second wave’s coded picks for both engines with their matches, and evidence routes for the ChatGPT repeat, and the full text of the 200 second-wave answers. Anyone can recode the answers from the text files. The methodology, the second wave’s plan with its hash, the coding rules and the scripts are available on request.

Cite as: VOCTOS (2026). ChatGPT backs three in four local business picks in Egypt with map data: 300 answers, two engines. Captured 13 and 14 October 2026. https://www.voctos.com/research/chatgpt-local-recommendations-egypt/

Revised on 14 October 2026. First revision: intervals resample whole answers, cautions attached to recommendations were coded, one missed review count was corrected, and a corrections policy was added. Second revision, the same day: a pre-registered second wave added a ChatGPT repeat and Google AI Mode on the same 100 questions. No first-run recommendation or route changed.

If you want to know where your own business stands in these answers, our تدقيق البحث بالذكاء الاصطناعي runs the questions your customers ask and shows what ChatGPT says. The work that follows usually sits with SEO المحلي for the listing and تحسين الظهور في ChatGPT for the pages it cites.