Nixeny Dijital - Mersin / AI Readiness 2026
The robots.txt file of 3,395 business websites in Mersin was fetched, and 3,382 home pages were examined for machine readability. The share of sites where an assistant can answer “where is this business, when is it open, what is its number” straight from the site is 4.5%. One question: what are Mersin’s businesses telling the machine?
What these numbers measure and what they do not
No AI model was asked a single question in this study. What assistants actually cite was not measured. The only thing measured is whether the conditions an assistant needs to produce an answer from this site are met. The sentence “AI cannot find this business” does not follow from this data.
Blocking AI crawlers is a legitimate decision, and it is counted in this report as a decision rather than as an error. The 3,021 sites that have written no rule at all and the 331 that deliberately block are never merged in any table: the first has not had the conversation, the second has had it and decided.
Content and structured-data measurements were taken from the home page alone. Schema, service lists or FAQs on inner pages are not seen. That means every schema and content rate here is a FLOOR: the real rate can be higher, not lower.
Bot names and policies change within months; a bot that does not exist today could top the list next year. The rates here are meaningful only together with the 19-bot list beside them, and cannot be carried to another date.
Glossary · every term explained once
- AI visibility
- The subject of this report: how many of the conditions an AI assistant needs to produce an answer from a business website are met. Not a quality judgement but a measure of readability.
- Base
- This report uses three main bases: 3,395 sites whose robots.txt was fetched, 3,382 sites with a readable home page, and 4,167 audited domains. Narrower bases appear inside sections (the llms.txt generators are over 352 files). Every rate names which one it is over.
- robots.txt
- A plain-text file at the root of a site saying which bot may go where. It is voluntary, not binding: bots choose to obey. In this study only the root path ("/") was evaluated for each bot.
- Training bot / search bot
- A training bot (GPTBot, ClaudeBot, Google-Extended) collects content to train a model. A search bot (OAI-SearchBot, PerplexityBot) finds and cites a source when a question is asked. Blocking both with the same rule also closes off the chance of being cited.
- No rule ≠ blocked
- If robots.txt carries no rule for a bot, the default behaviour is to allow access. That is why sites that have "written no rule" and sites that have "deliberately blocked" are never put in the same bucket anywhere in this report.
- Structured data (JSON-LD)
- Information embedded in the page in a form a machine can read directly: name, address, telephone, opening hours. Invisible to a person. Only JSON-LD was read in this study; microdata and RDFa are out of scope.
- LocalBusiness schema
- The kind of structured data that says "this is a local business". It is the standard way for a site to introduce itself to a machine as a business; without it an assistant has to guess what the page is from the text.
- Name-address-phone triple
- All three of the business's name, street address and telephone present in the structured data. If one of the three is missing, an assistant either withholds the answer or takes it from somewhere else — usually a directory.
- llms.txt
- A file proposed for summarising a site's content for AI assistants. It is still a draft proposal, not a standard. Its absence is therefore reported as a measure of adoption rather than as a shortfall.
- Answerability
- All four of the conditions an assistant needs to answer "where is this business, when is it open, what is its number" straight from the site: search not blocked, name-address-phone complete, opening hours present, and content visible without running JavaScript.
- JavaScript dependence
- Fewer than 120 words of text in the static HTML while the page the browser renders shows more than 2.5 times as much. These sites' home pages look empty to a crawler that does not run JavaScript.
- Visibility score
- A weighted 0-100 combination of six components: access permission (25), structured data (25), content depth (20), technical baseline (15), answer readiness (10), llms.txt (5). A component that cannot be measured is dropped from both the numerator and the denominator; below 60% coverage no score is published.
Four sites in a hundred can answer a question
For an AI assistant to answer the most ordinary question about a business — where it is, when it is open, what its number is — straight from that business’s own site, four things are needed: robots.txt must not be blocking search, the name-address-phone triple must be complete in the structured data, opening hours must be there too, and the content must be visible without running JavaScript.
The number of sites meeting all four is 152 — 4.5%. Those sites represent 170 businesses. For the rest, an assistant either says nothing or takes the answer from somewhere else, usually a directory.
Each bar is the number of sites failing that one condition. A site can fail three at once, so the rows do not sum and cannot be read as a partition; the ordering is by how often the condition bites.
89.0% of sites have written not one rule for AI
robots.txt is the file where a site says which bot it lets in where. 2,227 sites (65.6%) have one; but the number carrying a single rule inside it for AI crawlers is only 374. For the remaining 3,021 sites the subject has not come up yet.
Writing no rule is not blocking: if robots.txt carries no provision for a bot, the default behaviour is to allow access. That is why these two rows are never merged anywhere in this report. One has not had the conversation; the other has had it and decided.
Seven readings, from the same file and over the same base. The “at least one rule defined” row is the sum of the permission and refusal decisions; the two rows under it split that sum. The Googlebot row is the control.
It is possible to see the state of the sites as a single partition. The four buckets below cover all 3,395 sites and sum to the base — but the “closed” bucket is not the bottom of a ranking, it sits outside the other three. A site there has answered the question; a site in “invisible” has never met it.
Each of the hundred squares stands for about 34 sites. The buckets are mutually exclusive and sum to the base.
AI crawlers are blocked 25 times as often as search engines
The only way to tell whether the blocking is a general closed door or a distinction specific to AI is to look at what the same sites do to classic search bots. 331 sites block at least one AI crawler at the root path; the number blocking Googlebot and Bingbot is 13. The difference is 25.5×.
What is being refused, then, is not crawling but a particular use of what gets crawled. The same sites let search engines in.
Four groups, one base. The control group is the classic search bots: Googlebot and Bingbot. The “named” column counts the cases where the site writes that group's name in robots.txt — naming is not blocking, a bot can be named in order to be allowed.
Read bot by bot, the blocking turns out to be a list job: the block rates for training bots sit very close together, because the same sites name them in a single block. On a few rows the blocked count exceeds the mentioned count — there the rule is written by the wildcard (*) group and the bot’s own name never appears in the file.
Each row is a user-agent token. The colour shows the bot's job: red training, dark blue search, light blue fetch-on-user-request, grey classic search (control).
| Bot | Job | Named | Blocking | Rate |
|---|---|---|---|---|
| GPTBot (OpenAI, training) | training | 346 | 312 | 9.2% |
| OAI-SearchBot (ChatGPT search) | search | 62 | 40 | 1.2% |
| ChatGPT-User (on user request) | user request | 160 | 135 | 4.0% |
| ClaudeBot (Anthropic, training) | training | 336 | 308 | 9.1% |
| Claude-User (on user request) | user request | 40 | 40 | 1.2% |
| Claude-SearchBot (Claude search) | search | 43 | 41 | 1.2% |
| PerplexityBot (search index) | search | 171 | 131 | 3.9% |
| Perplexity-User (on user request) | user request | 32 | 38 | 1.1% |
| Google-Extended (Gemini training) | training | 334 | 305 | 9.0% |
| Applebot-Extended (Apple Intelligence) | training | 316 | 305 | 9.0% |
| meta-externalagent (Meta AI) | training | 214 | 218 | 6.4% |
| Bytespider (ByteDance) | training | 309 | 313 | 9.2% |
| CCBot (Common Crawl) | training | 320 | 318 | 9.4% |
| Amazonbot | training | 303 | 308 | 9.1% |
| cohere-ai | training | 44 | 45 | 1.3% |
| YouBot (You.com) | search | 127 | 131 | 3.9% |
| Diffbot | training | 125 | 136 | 4.0% |
| Googlebot (classic search) | classic search | 68 | 11 | 0.3% |
| Bingbot (classic search + Copilot) | classic search | 40 | 13 | 0.4% |
352 sites have an llms.txt; a plugin wrote half of them
llms.txt is a file proposed for summarising a site’s content for AI assistants. It is not a standard yet but a draft proposal — so its absence should be read as a measure of adoption rather than as a shortfall. 352 sites (10.4%) have one; the more detailed llms-full.txt is on only 30.
The real finding is not the number itself but who wrote it. 183 of these 352 files (52.0%) were generated automatically by a plugin or a platform. So this number measures not the businesses’ decision about AI but the release notes of the tool they use.
Two different bases, stacked. The rates in the lower block are the distribution of the files that exist, not of the site total: read against 3,395 every row comes out an order of magnitude wrong.
Sites a machine can read opening hours from: 5.7 percent
Structured data is the information embedded in a page that only a machine sees: name, address, telephone, opening hours. The share of sites with any JSON-LD block on their home page is 33.9%. But the number falls at every step below that: 10.7% introduce themselves as a business, 13.5% have the complete name-address-phone triple, and 5.7% give opening hours.
That fall is the report’s subject. The one field an assistant needs to answer “is it open now” exists on one site in twenty.
A narrowing sequence: any JSON-LD at all, then the business types, then the three facts an assistant actually needs to answer the question.
A completeness percentage over eight fields. The median is zero: more than half the sites carry not one countable field, so the distribution itself is given rather than an average.
The types an assistant can quote directly. FAQPage and Question carry a ready answer to a question; Offer, Service and Product say what a service is and what it costs.
155 sites’ home pages look empty to a crawler
Where there is no structured data, the text itself is what remains. The median home page carries 382 words; a quarter of them are under 176. For most sites there is something to read: 81.7% of sites are in the “adequate” or “rich” bucket.
The problem is at the extremes. 155 sites’ content arrives only when JavaScript runs — which means an AI crawler that does not run JavaScript sees that home page empty. A further 150 sites carry under 120 words on their home page.
Four buckets by the word count of the static text on the home page. The buckets are mutually exclusive and sum to the base.
The six marks an assistant looks for in the text when there is no structured data. A telephone link (tel:) is machine-readable without any schema at all; it is the only row above half.
These two rows are not the same sites and do not sum: the first counts sites that load their text afterwards, the second sites with no text on the home page at all. The result is the same either way: a bot that does not run the page like a browser finds nothing there to read.
The median site scores zero on two of six components
Six components give a weighted 0-100 score: access permission (25), structured data (25), content depth (20), technical baseline (15), answer readiness (10) and llms.txt (5). Median 53.8, quartiles between 44.5 and 63.7.
The score is not a quality judgement but a measure of visibility: a low score does not mean a bad business, it means less for a machine to read. Looking at the components shows where the distribution comes from — the median site takes full marks on access permission and zero on two components.
Five buckets summing to the base. The distribution gathers in the middle: half the sites are in the 40-59 band.
Each of the hundred squares stands for about 34 sites.
The component's weight is written beside it. Two components have a median of zero: that is not a gap in the data but the finding itself — more than half the sites have neither structured data nor an llms.txt.
What decides this is not the sector, it is the platform
The gap between sectors is narrow. The medians of 30 sectors are squeezed between 33.8 and 64.4, and their median is 52.9. The gap between platforms is wider: 68.9 on Wix, 48.8 on custom builds.
The row that shows it most plainly is the LocalBusiness schema: 53.7% of Wix sites introduce themselves to a machine as a business, against 8.9% of custom builds. What creates the difference between a lawyer’s site and a car workshop’s site is not law and car repair; it is the tool the two sites were built with.
The two bases on a row are not the same, and should not be: the median score is over all probed sites, while the LocalBusiness rate is over those whose home page could be read.
| Platform | Sites | Median score | LocalBusiness | Name-addr.-tel. | Answerable |
|---|---|---|---|---|---|
| Wix | 121 | 68.9 | 53.7% | 43.8% | 1.7% |
| Ticimaxcomparison | 48 | 66.9 | 4.2% | 81.3% | 4.2% |
| Next.js | 146 | 61.4 | 39.0% | 39.0% | 19.2% |
| WordPress | 559 | 59.4 | 7.7% | 8.8% | 3.8% |
| WooCommerce | 386 | 58.8 | 5.2% | 5.7% | 2.3% |
| IdeaSoftcomparison | 37 | 56.3 | 0.0% | 5.4% | 0.0% |
| Custom build | 1,912 | 48.8 | 8.9% | 11.2% | 4.6% |
Each tick is one sector's median score. The tick height carries no information; it varies on a three-step cycle only to keep values that land side by side countable. The red line is the median of the sector medians.
The finding of this figure is its flatness. Where every other measure in the series runs several times over between districts, AI visibility does not travel with geography.
This study asked no AI anything
Every number in this report measures a readiness: how much of what an assistant would need to produce an answer from this site is in place. None of them measures what an assistant actually does.
Not a single question was put to any model in this study. The sentence “opening hours are not in the structured data” is true and verifiable; the sentence “AI cannot find this business” does not follow from this data. Assistants also produce answers from directories, from map services and from sources this study never looked at. That distance is one this report did not set out to close.
In the same way, 331 sites blocking AI crawlers is not a defect. A business may not want its content used in model training, and this is exactly how that is said. This report counts that decision; it does not judge it.
- That local business sites are not configured to answer AI assistants — the share meeting all four conditions is 4.5%.
- That AI crawlers are not yet on local businesses' agenda: 89.0% of sites have not written a single rule for these bots.
- That the blocking is a distinction specific to AI: the same sites let classic search bots in, a difference of over 25 times.
- That the difference comes largely from the defaults of the platform in use: the median gap between platforms is larger than the gap between sectors.
- That the 155 sites whose content arrives only with JavaScript look empty to crawlers that do not run it.
- “AI cannot find these businesses” — no model was asked anything; conditions were measured, not outcomes.
- “That site is doing the wrong thing by blocking AI” — blocking is a legitimate choice and is counted here as a decision.
- “X% of sites have no schema” — the accurate phrasing is “none was detected on the home page”; inner pages and non-JSON-LD formats were not seen.
- That the bot rates are durable — bot names and policies change within months.
- That sites publishing an llms.txt have made a decision about AI — more than half the files were generated automatically by a plugin.
Limitations
- A single cross-section.
- robots.txt and llms.txt were measured as they stood at the time of the fetch. A site may have changed its rule the next day; this study shows that moment, not the change.
- Conditions were measured, not outcomes.
- What assistants actually cite was not measured. Being allowed does not guarantee being cited; being blocked does not mean being invisible.
- Bots are not obliged to obey robots.txt.
- The file is a voluntary agreement. Some bots do not obey it. So “blocked” does not mean “did not enter”, it means “asked not to enter”.
- Only JSON-LD was read.
- Structured data was assessed through JSON-LD blocks alone. Microdata and RDFa are out of scope; sites using those formats appear here as “no schema”.
- Content measurements are limited to the home page.
- Depth on blog, service and contact pages is not seen. A site with a rich inner structure can look low because of a thin home page.
- The sector assignment is a heuristic.
- The sector is derived from the search term that found the business. For businesses with several lines of work another assignment could be defended; the sector rankings depend on this rule.
- District differences carry a small sample.
- In three districts the number of audited sites is under 30. Those rows stay in the table but do not enter the comparison and are not taken into a headline.
Citation rules
These six rules are not a matter of style. Framing a number wrongly in this report turns a choice into a defect, or a measure of readiness into a prophecy.
- 01Blocking is a choice, not an error; "wrote no rule" and "blocked" are reported separately.
- 02The score is a measure of visibility, not a quality judgement; a low score does not mean a bad business.
- 03Because only the home page was measured, the content and schema rates are a floor.
- 04The bot list is date-stamped; on another date there will be other bots.
- 05Do not say "AI cannot find this business"; say "the conditions it would need to answer from this site are missing".
- 06llms.txt is still a draft proposal; its absence is presented as a measure of adoption, not as a shortfall.
The method, briefly
- 01UniverseThe data set of the Mersin / State of Digital 2026 study: 4,167 audited domains, 3,395 of them reachable. The business universe is the 19,615 local businesses visible as active on Google; behind these sites stand 3,928 businesses.
- 02The fetch, and its courtesy rulesAt most three GETs per domain: /robots.txt, /llms.txt, /llms-full.txt. A self-identifying User-Agent, low concurrency and a wait between requests were used. robots.txt was read first, and no llms.txt request was made to sites blocking this research bot. The results were written to a database, so re-running produces no new requests.
- 03How robots.txt was evaluatedOnly the root path ("/") was evaluated for each bot. Matching is case-insensitive, the longest path rule wins, allow takes precedence on a tie, and an exact token match overrides the wildcard group. "No rule" and "blocked" were counted separately.
- 04Structured data and contentThe home-page HTML was read from the 2026·08·25 cache; no new request was made for this section. JSON-LD blocks were parsed and broken blocks counted separately. Content measurements were taken on the static HTML.
- 05The JavaScript testWhere the static HTML held under 120 words and the page Playwright saw was more than 2.5 times larger, the content was counted as JavaScript-dependent. This test could be run on 3,364 sites.
- 06Score and thresholdsA weighted 0-100 combination of six components; a component that cannot be measured is dropped from the numerator and the denominator, and below 60% coverage no score is published. 3,382 sites were scored, 13 could not be. Sectors entered the comparison only with 50 or more audited sites; 30 sectors passed that threshold.
| Audited domains | 4,167 |
|---|---|
| Sites whose robots.txt was fetched | 3,395 |
| Sites with a readable home page | 3,382 |
| Scorable sites | 3,382 |
| Businesses behind the sites | 3,928 |
Sources
- 01robots.txt / llms.txt fetch (this study)The series' only new field measurement. At most three GET requests were made per domain: /robots.txt, /llms.txt and /llms-full.txt. No other page was requested. The fetch window was 2026·08·27 16:28 → 16:44 UTC. robots.txt was read first; no llms.txt request was made to sites blocking this research bot at the root path.
- 02Nixeny Dijital · Mersin / State of Digital 2026The same universe, the same set of businesses. The 4,167 domains, 3,395 reachable sites and 3,382 home-page caches this report uses come from that study; the home-page HTML was taken on 2026·08·25. Structured-data and content measurements were made on that cache; no request was sent again.nixeny.com
- 03The bot list19 user-agent tokens: bots from OpenAI, Anthropic, Perplexity, Google, Apple, Meta, ByteDance, Common Crawl, Amazon, Cohere, You.com and Diffbot, plus Googlebot and Bingbot as a control group. The list is date-stamped and set out openly in `config/ai-readiness.json`; on another date there will be other bots.
- 04robots.txt interpretation rulesOnly the root path is evaluated for each bot. Matching is case-insensitive, the longest path rule wins, allow takes precedence on a tie, and an exact token match overrides the wildcard (*) group.www.rfc-editor.org
- 05schema.org · LocalBusinessThe definition of the structured-data fields. The fields sought in this study: name, address, telephone, openingHours / openingHoursSpecification, geo, priceRange and aggregateRating.schema.org
- 06The llms.txt proposalThe draft defining the file's format. It is a proposal, not a standard; this report measures adoption, not compliance.llmstxt.org
The numbers in this report were not compared with a measurement of another province or country. A “world average” produced with a different bot list, on a different date and over a different base looks, when placed side by side, as though both numbers measure the same thing. What is shared with Mersin / State of Digital 2026 is the universe and the date of the home-page cache; the robots.txt fetch is specific to this study and was made two days later.
Colophon
Conducted and published by Nixeny Dijital, Mersin. The fetch, the scoring and the report design are Nixeny’s own. The universe and the site list are shared with the Mersin / State of Digital 2026 study. At most three requests were sent per domain for this report: robots.txt, llms.txt and llms-full.txt.
If a business wants to know its own site’s row in this study, writing in the domain name is enough. The method, the bot list and the verification logs are also shared on request: info@nixeny.com