casestudy #4 · BINARYX × 009
×238 more visits from AI assistants
since april.
April 2026: a tokenized real estate platform appears in 3% of AI answers about its category, and the site it has just migrated cannot be read, verified or cited by the engines. Below is a GEO programme of five months in four languages: every SEO fix in it was made because it changed what ChatGPT, Perplexity and Gemini could read, each change measured against a control group, and one finding that changed the plan.
- client
- Binaryx tokenized real estate platform: entry from $500, 2,000+ investors, properties in Bali, Turkey, Montenegro, Dubai and more
- niche
- tokenized real estate, RWA rivals in AI answers: Fundrise, RealT, Lofty, Arrived, RealtyMogul, Roofstock, Ark7, CrowdStreet, Yieldstreet, Stake
- website
- binaryx.com EN (US, UK), UA, RU, FR; blog of 464 articles; Webflow to Astro migration in April 2026, Keystatic CMS
- programme
- 16 apr 2026 → still running weekly snapshot every saturday: 35 keywords in 2 geos, GSC, GA4, AI panel of 50 prompts
- status
- named, with the client’s approval brand, dates and search-console figures published with Binaryx’s consent
vocabularywhat is what on this page
- GEO
- Generative engine optimization: the work of being read, verified and cited by AI engines. On this page every SEO task exists because an engine reads the same page, the same index or the same redirect.
- control group
- Pages or languages that did not get the change, measured over the same window. Every headline figure on this page is stated next to its control.
- T+28, T+90
- The two fixed re-measurement points after a change ships: 28 and 90 days. Nothing is called a result before T+28.
- hypothesis ledger
- A record written before each deploy: what changes, what is expected to move, and the rule for rolling it back. 87 entries over five months.
- share of voice
- How often the brand is named across a fixed panel of 50 category prompts on two engines, ranked against the rivals named in the same answers.
- AI-assistant visits
- Sessions that GA4 attributes to ChatGPT, Gemini, Perplexity, Claude and Copilot. Counted in the site’s own analytics, never estimated.
- IndexNow
- The protocol that pushes URLs into the Bing index directly. It is the reason the ChatGPT chapter below exists.
where it stands now:
french blog, 112 articles rewrittenfrom zero
0→ 150clicks a month
150 clicks a month at 90 days from a blog that had none; the untouched EN, RU and UA blogs moved +7% to +30% in the same window
share of voice in AI answers+9 places
13th→ 4thof 14
first panel run, 17 may: behind everyone but one. latest run: above Stake, Ark7, Yieldstreet, CrowdStreet, Roofstock
brand query in bing#1
>20→ #1with sitelinks
baseline: outside the top 20, two clone sites instead. latest: first, blog indexed, clones at #5 and #8
visits from AI assistants×238
3→ 713a month
baseline → latest; ChatGPT first, then Gemini, Perplexity, Claude and Copilot
what was actually broken here?
A migration had just shipped, and the map behind the site had not. Everything below is the locked baseline of 16 april 2026; every “latest” figure on this page is measured against it, never against a better later run.
| at the baseline | what it meant |
|---|
| hundreds of URLs returning errors after the Webflow to Astro move | years of links and rankings pointing at nothing |
old /nl and /uk locales and ~1,000 tag archives in the index | the crawl budget spent on pages nobody should see |
| empty meta fields; CMS admin showing 115 of 476 posts | the team could not even see its own blog |
| a scam clone ranking for the brand in two countries | buyers who searched the name could land on a fraud |
| named in 3% of AI answers about the category, 13th of 14 in share of voice | Fundrise, RealT, Lofty, Arrived, RealtyMogul, Roofstock, Ark7, CrowdStreet, Yieldstreet and Stake in the answers; Binaryx not |
| brand query in Bing outside the top 20, blog not in the Bing index | invisible to ChatGPT search, which reads Bing; the reason took until july to find |
the programme · date by date
we ship. the engines react.
← measured · what the data said
shipped → · what we changed
16 apr 2026measured
the baseline: a fresh migration and a broken map.
The site had just moved from Webflow to Astro. Organic clicks were down, hundreds of URLs returned errors, the old /nl and /uk locales and about a thousand tag archives sat in the index, meta fields were empty, the CMS admin showed 115 of 476 posts, and a scam clone ranked for the brand in two countries. The brand was named in 3% of AI answers about its category and sat 13th of 14 in share of voice; in Bing it was outside the top 20.
✕ fundrise, realt, lofty, arrived, realtymogul, roofstock, ark7, crowdstreet, yieldstreet, stake: named. binaryx: not. every fix below exists to change this line.
act one · 18 to 30 aprshipped
the wreckage first: an engine cannot cite a page that returns an error.
31 blog redirects and the meta of the main pages in four languages in the first three days. Then the taxonomy: 7 hubs and 22 tags across 468 articles, Webflow leftovers cleaned from 504 files, redirects for /uk, /nl and /tag (157 URLs, about 41K impressions a year). Person schema and author pages for 14 authors in 4 languages. On 30 april alone 12 pull requests: 142 redirects, 16 legal pages in 4 languages, sitemap from 533 to 636 URLs, the CMS admin repaired. Complaints on the clone filed through six channels.
act two · may to julshipped
pages the engines can quote: content in four languages, each batch against a control.
Bottom-of-funnel pages the buyers actually search for, arrived-homes-alternatives and fundrise-alternatives, in four languages. Meta rewritten in 174 articles, 3,336 diacritics restored in 80 French articles, disavow file to 471 domains. 112 French articles rewritten in full, the other three blogs left untouched as the control. 548 blog URLs inspected, about 1,050 tag archives set to noindex, 148 broken pages closed, 301s on 188 phantom URLs, FAQPage schema across the blog, 33 investor stories marked up as VideoObject, 41 deep links from articles to properties. A demand map of 613 queries and a ledger of 87 hypotheses, each written before deploy with its expected effect and rollback rule. Seven waves, 66 Ukrainian articles in nine days. Stale property statuses corrected in 87 articles so the engines would stop quoting listings that no longer existed.
act three · 11 julshipped
the finding: ChatGPT ignored the site because Bing had never seen the blog.
The question was why ChatGPT search returned nothing from a site Perplexity already quoted. The answer was one layer below any GEO checklist: ChatGPT search reads the Bing index, and the blog was not in it. The fix shipped the same day, IndexNow on 997 URLs, llms.txt checked line by line against the site. Two more clone domains surfaced on the way; by 1 august their cloaking was proven and reported: Bing was shown a copy of the site, people and Google got a redirect. AI-bot impressions were filtered out of the monitoring so the numbers would stop flattering themselves.
16 sep 2026measured
the latest run: every row of the table above.
French blog clicks 0 → 150 a month against controls at +7% to +30%. Ukrainian “tokenized real estate” #4 → #1. Share of voice in AI answers 13th → 4th of 14. Brand in Bing from outside the top 20 to #1 with sitelinks, clones at #5 and #8. Visits from AI assistants 3 → 713 a month, the deepest-reading channel on the site. Twenty-five weekly reports on file, 70+ pull requests merged.
the other half · what engines say when you name the platform
why ChatGPT ignored us: bing.
The site had schema, an llms.txt and pages Perplexity already quoted. ChatGPT search still returned nothing from it. ChatGPT search reads the Bing index; the blog was not in Bing, and the brand query there returned two clone sites instead of the platform. Nothing on a page can fix an index the page is not in.
1baseline · bing, brand queryOutside the top 20. Blog absent from the index. Two clone domains ranking instead, one of them showing Bing a copy of the site and everyone else a redirect.
→
2the fixIndexNow on 997 URLs, llms.txt checked against the facts on the site, clone tracking in the weekly monitor, cloaking documented and reported.
→
3latest · bing, brand query#1 with sitelinks, blog indexed, clones at #5 and #8. ChatGPT is the largest AI source of visits on the site.
what this page does not claim: that the brand “dominates ChatGPT”. The API-based panel queries a model, not the live product. The claim here is the one the analytics support: visits from ChatGPT, and the Bing index they depend on.
the sources · what the engines cite now, against the baseline
the engine quotes the platform’s own page, word for word.
At the baseline the category answers were assembled from the rivals’ pages and from directories. In the latest run Perplexity cites the platform’s own comparison article first and repeats its facts verbatim.
perplexity · category prompt, latest logged run
What are the alternatives to Arrived Homes for fractional real estate investing?
… Binaryx offers tokenized real estate with 38 properties across 4 countries, from $500, targeting 8 to 12% yields … source 1: binaryx.com/blog/arrived-homes-alternatives
reconstructed from the logged answer. Other pages Perplexity cites: buying-tokenized-real-estate, the Montenegro and Bali market articles. Google’s AI Overview cites the site for “best rwa tokens” and “metamask alternatives”. The page it quotes first is the bottom-of-funnel article published in act two.
the search side · the SEO that fed the engines, with controls
the SEO we did because the engines read the same pages.
None of this was done for rankings alone. Perplexity, ChatGPT search and Google’s AI Overview assemble their answers from indexed, ranked, readable pages, so the classic SEO work was the way in: redirects so the engines stop landing on errors, indexation so the pages exist in the index the engines read, rewrites so there is something worth quoting. Two dates per row, the locked baseline and the latest logged run, every change next to the pages that did not get it. Source: Google Search Console, GA4 and the AI panel, pulled 16 september 2026.
| measure | baseline | latest | control or note |
|---|
| french blog clicks a month (112 articles rewritten) | 0 | 150 | controls EN, RU, UA +7% to +30% |
| “tokenized real estate” in ukrainian | #4 | #1 | five more RU and UA tokenization queries: #3 to #9 → #1 to #2 |
| “arrived homes alternatives”, US | – | #2–3 | page did not exist at baseline; UA, RU and FR versions at #1 |
| “blackrock tokenizing assets”, US and UK | #15 / #16 | #5 / #5 | |
| blog clicks a month | 345 | 557 | +61%, of which 150 from the French rewrite; site-wide clicks −17% in the same window, see the last question in the FAQ |
| non-branded queries with clicks | 38 | 58 | +53% |
| blog pages indexed | 76.8% | 81.8% | 548 URLs inspected, about 1,050 archives removed from the index |
| brand query in bing | >20 | #1 | with sitelinks; clone sites at #5 and #8 |
| share of voice, 50 prompts on 2 engines | 13th of 14 | 4th | above Stake, Ark7, Yieldstreet, CrowdStreet, Roofstock |
| visits from AI assistants a month | 3 | 713 | ChatGPT first, then Gemini, Perplexity, Claude, Copilot |
| time on site per AI-assistant visit | – | 190 s | 44.5% engaged; paid social 71 s at 23.7%; Gemini visitors 394 s, ChatGPT 245 s |
what workedrewriting pages that already existed, and bottom-of-funnel pages that intercept a rival’s name. New articles for competitive head terms did not: three of three without a click.
the point of all of it · what the shelf produces
the answer turns into a visit that reads.
Being cited is not the point. Being cited is the moment a buyer who was going to open a rival’s page opens yours instead. On this site that moment is measured in its own analytics.
1a buyer asks“What are the alternatives to Arrived Homes?” No brand in the question. They do not know the platform exists.
→
2the engine answersBinaryx, with its own comparison page cited as the first source and its facts repeated: 38 properties, 4 countries, from $500.
→
3the visitor arrivesFrom 3 to 713 visits a month from AI assistants, 190 seconds and 1.6 pages each: the channel that reads the most.
entry pages from AI assistants
fractional real estate
44 visits
turkey market page
41 visits
before you ask
questions about this case.
What exactly did 009 do for Binaryx?
A GEO programme: making the platform readable, verifiable and citable by AI engines. The SEO tasks inside it were chosen because each one changed what the engines could read. Five months after the migration from Webflow to Astro: about 700 redirects and the clean-up of hundreds of broken URLs, taxonomy and indexation of a 464-post blog in four languages, Person, FAQPage and VideoObject schema, a full rewrite of 112 French articles and 66 Ukrainian ones, bottom-of-funnel comparison pages, IndexNow onto 997 URLs after finding the blog missing from Bing, and a weekly measurement of positions, AI answers and AI-assistant visits. Over 70 pull requests were merged into the site.
Why is there so much SEO in a GEO case?
Because the engines do not have a separate internet. Perplexity cites indexed pages, ChatGPT search reads the Bing index, Google's AI Overview quotes ranked results. A site that returns errors, hides its blog from the index or has nothing quotable on the page is invisible to all of them at once. Every redirect, every rewrite and every indexation pass on this page was done for that reason, and the AI-answer measurements are where the result shows.
Why is every result shown against a control group?
Because a number without a control proves nothing on a site whose overall traffic was falling. Each change was logged in a hypothesis ledger before it shipped, with the expected effect and a rollback rule, and measured at 90 days against pages that did not get the change. The French rewrite is the cleanest example: from zero to 150 clicks a month while the untouched English, Russian and Ukrainian blogs moved +7% to +30%.
How do you rank in ChatGPT?
For this site the answer was Bing. ChatGPT search reads the Bing index, and the blog was not in it, so no amount of llms.txt or schema could help. Submitting 997 URLs through IndexNow and fixing what Bing saw put the brand at #1 in Bing with sitelinks by mid-August, from outside the top 20 in July. The visits from ChatGPT are counted in the site's analytics, not claimed from a screenshot.
Did the whole site's traffic grow?
No, and the case does not claim it. Site-wide clicks fell during the period because branded demand dropped after the client paused an awareness channel in late May. The results here are for the blog and for non-branded queries, where the work happened and where the controls live.
Did Binaryx agree to be named?
Yes. Client names stay under NDA by default; Binaryx approved a named case with its search-console figures. The internal hypothesis ledger and weekly reports stay with the client.
Program: 16 april 2026 → ongoing; the last three months ran as weekly sprints, with a snapshot every saturday. Published sep 17, 2026. Every figure here is a snapshot of the dated run named beside it; nothing on this page updates by itself. The baseline is locked at 16 april 2026 and every “latest” figure is the logged run of 16 september 2026, never a better run in between. Search-side figures come from Google Search Console, visits from the site’s own GA4, share of voice from a panel of 50 prompts on two engines, each answer logged on the day it says. One answer from a generative engine is noise rather than a measurement. Answer panels are illustrative renderings of logged runs. Brand, dates and search-console figures published with Binaryx’s consent.