Playbook
How to be cited by SI assistants
SI assistants such as ChatGPT, Gemini, Perplexity and Google AI Overviews no longer hand a buyer a page of links. They write an answer, name one or two companies and show a handful of sources. This playbook collects what Kaan Gülten has said in his talks about getting into that answer, in seven steps. Every step links to the slide it comes from; the talks were given in Turkish.
“‘Being in first place’ now means ‘being mentioned inside the answer’.”
The Change in Search, slide 7, Ankara E-Export Summit 2025 · translated from Turkish
By Kaan Gülten · Based on talks given in Turkish · Published · 7 min read
From clicks to citations
In classic search, the goal was to reach the top of the results and win the click. In an SI answer, the goal is to appear inside the answer, for the right reason. Kaan Gülten describes the shift as a change of currency:
| Click economy | Citation economy | |
|---|---|---|
| Goal | Reach the top of the results | Appear in the answer with the right reason |
| Unit | Keyword and page | The buyer's question and need |
| Success | Ranking, click, session | Mention, recommendation, citation |
| Content | Long and keyword-heavy | Information that stands on its own and can be verified |
| Authority | Backlink profile | Consistent information and independent sources |
Classic search has not ended; the goal has grown to cover a wider evaluation journey. What has changed is the cost of being third. In a page of links, third place was still considered. In an SI answer there is no list: one or two suppliers are recommended, and the third does not exist.
Step 1. Let the right crawlers in
Three different kinds of bot come to a website, and they come for different reasons. A training bot reads the site to train a model; blocking it is a company decision. A search bot comes so that the answer can cite you; block it and you are left out of the story. A user-request bot arrives because a buyer is asking about you at that moment; block it and you turn the customer away at the door. A security setting can be sending away the second and third without anyone noticing.
The starting policy Kaan Gülten proposes:
- Allow the search and user-request bots: OAI-SearchBot and ChatGPT-User (OpenAI), Claude-SearchBot and Claude-User (Anthropic), PerplexityBot (Perplexity).
- Always allow Googlebot and Bingbot: they serve search, Google AI Overviews and Copilot.
- Treat the training bots (GPTBot, ClaudeBot, Google-Extended) as a company choice.
robots.txt is not a security tool: private documents are protected by authentication, and firewall and CDN rules are checked separately. Then check that the critical information is in the first HTML. If a technical table only loads later through JavaScript, some systems may never see it. Asking a model “can you read my site?” is not a technical test; opening the page source and searching for a product code is.
Step 2. Write information that stands on its own
Superintelligence describes a company correctly only when its information is clear, numerical, sourced and current. Google's SI answers break a page into parts and summarise it, so each part has to make sense on its own: what it is, and under which condition it applies. Google's systems also dislike ambiguity, which means the company name and the claim belong in the same sentence. The talk gives an example with a placeholder name:
“So-and-so Textiles is an İzmir-based manufacturer that exports 5 million metres of denim a year to Europe.”
The New Digital Customer, slide 9, IGEXX 2026 · translated from Turkish (“Falanca” is the Turkish placeholder name)
Content that SI systems can use is structured (Schema.org, JSON-LD), gives a clear and direct answer, is public and crawlable, and is supported by authoritative sources.
“SI rewards not the quality of content, but how easy it is to understand.”
The Change in Search, slide 10, Ankara E-Export Summit 2025 · translated from Turkish
Step 3. Publish your own numbers
A company holds data nobody else has. Publishing it as statistics gives answer engines something to cite that can only come from you. Every value should rest on a document, a test or a record; if there is no proof, it is not written.
“You have data about your own company. You publish its statistics. That becomes information ChatGPT accepts as a reference more often. What raises visibility is how you present your own data to them.”
The Change in Search, slide 10, Ankara E-Export Summit 2025, from the recorded talk · translated from Turkish
Step 4. Be present where answers draw from
SI systems look less at what a brand says about itself and more at what others write about it. When building an answer they go to comparison and list sites, to discussion forums such as Reddit, to review platforms and to supplier platforms. For an exporter, appearing in the trade press of their sector, in German, Turkish or English, can be worth more than appearing in a large newspaper, and a factory manager's LinkedIn post can do more than the corporate brochure.
“They don't trust you; they trust the people who talk about you.”
The New Digital Customer, slide 10, IGEXX 2026 · translated from Turkish
Three actions from the talk: after every delivery, collect a concrete, sector-specific customer review; be present on international supplier platforms, on Wikipedia and in trade publications; have executives write on LinkedIn under their own names. The buyer who sees your name in an answer goes on to check the people behind the company, so the data on your site and the person on LinkedIn should say the same thing.
Step 5. Keep one version of the truth
Being more visible also makes wrong information more visible. The same product is usually described in several places: the web page, its structured data, the product feed or catalogue, and the quoting or ordering system. A human buyer who notices a difference asks. An agent does not ask; when it cannot tell which version is true, it moves on.
In the fictional case that runs through the UİB training, the product page said 500 g/m² and the feed said 450 g/m². The agent could not decide which to trust and left the towel out of the comparison. The rule that follows is simple: one code per product, the same code everywhere. Use the barcode (GTIN) if there is one, otherwise your own part number, and never invent a code.
Step 6. Correct wrong information at its source
A poor perception grows like a snowball. A few negative reviews enter an SI system; it starts to say “customer service is bad”; people who read that repeat it; the system reads those posts again.
“Instead of getting angry, you need to find the source that feeds this information and correct it.”
The New Digital Customer, slide 10, IGEXX 2026 · translated from Turkish
When you measure (step 7), a citation that carries wrong information does not count as correct information. Brantial, the company Kaan Gülten founded, has published a report on how outdated or wrong information about brands ends up in SI answers (AI Brand Misinformation Report, 2026; company research).
Step 7. Count mentions, recommendations and citations separately
Being mentioned, being recommended and being cited are three different results. Mentioned: your name appears in the answer. Recommended: the answer says you suit the buyer's need. Cited: your website is linked as a source under the answer. Ask the same questions every month, count the three separately, and never decide on a single screenshot.
For each answer, record: are we mentioned, are we recommended, is our site a source, where are we in the list, is the information about us correct, and who appears before us. A single answer is not a measurement; repeat the same question on different days and accounts.
A five-step diagnosis to run this week:
- Choose one target market, one product and one type of customer who buys it.
- Ask ChatGPT, Gemini and Perplexity the question your buyer would ask, without typing your brand name.
- Note whether your name appears, which competitors are there and which sites are cited.
- Compare the product's size, minimum order, lead time and certificate details on your website, your catalogue and the marketplace.
- Give every inconsistency an owner and a date. Ask the same questions again after 30 days.
Part of the effect leaves no trace: a buyer who saw your name in ChatGPT and visits two days later shows up in analytics as “direct traffic”. Add one question to your quote form, “Where did you first hear about us?”, and include “SI assistant” among the options.
What does not work
- Manufactured brand mentions, fake reviews and hidden advertising do not build trust.
- Unproven adjectives such as “leader”, “best” or “fastest” do not help. References, test results and certificates are used only if they are true and you have the right to publish them.
- Structured data does not guarantee visibility, and Google has no special schema for SI. Schema gives machines, in order, information that is already written on the page; it never carries information the page does not.
“There is no shortcut. I am not offering you a guarantee; I am proposing measurement and controlled experiments.”
Branding and Marketing in the Age of SI, slide 99, UİB Academy 2026 · translated from Turkish