Social Rocket
ChatGPT and other AI search tools are becoming a new discovery surface for brands. However, there is no guaranteed method to get recommended there: OpenAI explicitly states that search results are ranked based on several factors and that placement is not guaranteed.
Still, companies can build some concrete foundations. These include, above all, crawlable websites, original content based on your own experience and data, and pages that answer individual user questions clearly.
1. Technically ensure that ChatGPT can find the website
The most important technical point is surprisingly mundane: the page must be reachable for the search crawler. OpenAI names the OAI-SearchBot for this. If you accidentally block it in the robots.txt, in the CDN or through bot protection, you make it harder for your content to be included in ChatGPT search results.
OpenAI therefore recommends that publishers allow OAI-SearchBot. Tech teams should also check whether firewall, WAF or CDN rules block automated access. Allowing access only makes a website findable – it is not a ranking guarantee.
For marketing teams, a short technical check is worthwhile:
- Is OAI-SearchBot allowed in the robots.txt?
- Do important landing pages deliver usable content without a login or unnecessary JavaScript hurdles?
- Do Cloudflare or other protection systems block legitimate crawlers?
- Are the sitemap, canonicals and indexing rules properly maintained?
2. Publish content that cannot be easily replaced
The currently most important content lever is less "writing for AI" and more publishing something an AI cannot arbitrarily generate itself. In a recent Social Media Examiner interview, AI strategist Liron Segev therefore primarily recommends proprietary data, concrete experiences, case studies and specific results.
This is also relevant for social and creator marketing. A generic article such as "10 tips for better Reels" competes with thousands of similar pages. A post with your own campaign data, traceable methodology or a detailed creator case, by contrast, provides additional information that a search system can refer to.
This shift is already visible in creator marketing: briefings increasingly also consider visibility in AI search, not just reach on the platform itself.
3. Answer individual questions clearly and self-contained
For AI visibility, content does not have to sound artificially machine-like. However, a structure is helpful in which individual answers remain understandable without lengthy context.
In practice, this means:
- specific subheadings instead of vague chapter titles,
- the core answer early in the section,
- numbers always with source and context,
- clear terms instead of unnecessary marketing language,
- individual sections written so they are also understandable in isolation.
This is not a new SEO secret, but good information design. It helps people scan a page and at the same time makes relevant passages easier to find.
New retrieval observation: ChatGPT appears to select pages section by section
A recent finding from the GEO scene offers an unusually detailed look at the ongoing retrieval process. Metehan Yeşilyurt of Peec AI reports that on 8 September 2026 he observed, in ChatGPT's Server-Sent Events, an extensive debug representation of web search. OpenAI has so far not publicly documented this representation or confirmed the details derived from it. The following points are therefore observations by an external researcher, not an official ranking document.
In a test prompt about the "best AI visibility tools", ChatGPT reportedly ran five search rounds according to Yeşilyurt, formulated 18 different search queries, triggered 50 engine calls, collected 228 results and retrieved 223 URLs for further processing. Only 16 sources ultimately appeared in the final answer.
Even more interesting for publishers is the observed processing of individual pages. According to Yeşilyurt, page content is not passed to the answer model as full HTML, but first converted into a Markdown-like text representation. The observed format strongly resembles the Python tool `html2text`, though apparently not as an unchanged standard implementation. This rendered text is then split into blocks of roughly 170 words and evaluated. In his samples, usually only one to three selected blocks per source made it into the final context.
This would explain why self-contained sections can matter more than a perfect overall composition of the page. An important figure, limitation or core statement should ideally sit in the same section as the question it answers. Information that is only explained several screens later can more easily be separated out with chunk-based selection.
Yeşilyurt also reports that the observed stream carried, among other things:
- rewritten search queries and the search/retrieval engines used for each,
- scores for individual search results and individual text blocks,
- a decision on whether a result should be fully retrieved at all,
- crawl and publication data,
- a further sorting of domains before answer generation,
- and a fixed word budget for individual sources.
Metadata also appears to play a role. In Yeşilyurt's samples, raw `meta_tags` were transported with the results, while a separate `og_data` object remained empty. This does not yet show which meta tags actually influence ranking. However, it argues against the idea that ChatGPT processes only the visible body text of a page.
What this observation could mean for AI visibility
If the behaviour is confirmed by further tests, several practical consequences follow:
- One user question can generate many hidden search queries. Content should therefore not only match one exact phrasing, but sensibly also cover adjacent terms and decision questions.
- Being found is not the same as being cited. In the described test, more than 200 URLs were processed, but only 16 sources were visibly cited.
- Sections compete with each other. A good page can be relevant overall while only a single block wins or loses the selection process.
- Core statements should be understandable locally. Heading, statement, number, limitation and source should ideally sit close together.
- Keeping metadata clean remains sensible, even though no reliable weighting of individual tags can be derived from the observation.
The most important caveat: this is a snapshot of an observed production system. Retrieval pipelines, models and internal thresholds can change at any time. The data is therefore better suited to justifying robust content principles than to inventing new "170-word ranking hacks".
Existing content also does not have to disappear into the archive. For platform formats such as newsletters, a clean revision can create additional discoverability. Our guide on how to further optimise older LinkedIn newsletters fits here.
4. Use questions from support, sales and the community as a topic source
Instead of simply working through keyword lists, Segev recommends collecting real customer questions from support tickets, sales conversations and community comments. For brands this is especially useful, because these questions often sit closer to real decision problems than generic search terms.
A SaaS provider, for example, should not only explain what its product can do. Questions such as these are also more relevant: For which team size is it worthwhile? Which data is needed? Which integrations are missing? What does switching cost? Which mistakes occur in practice?
This creates a topic cluster along the actual customer journey rather than only around the product name.
5. Measure AI visibility – but not with a single metric
In its publisher FAQ, OpenAI explains that ChatGPT adds the parameter utm_source=chatgpt.com to referral links. This makes it easier to recognise incoming traffic from ChatGPT in analytics.
That alone is not enough as a metric. Brands should also observe which questions they appear for in AI answers at all, which pages are cited, and whether qualified visits or conversions result. The new retrieval observation additionally underlines that fetches and selected text blocks should be considered separately from visible citations, provided the corresponding measurement data is available at all.
As with social media, visibility without a suitable business goal has only limited significance. You can find a framework for this in our overview of social media goals and KPIs.
What marketers should take away
AI visibility is 2026 not a replacement for SEO, PR or social media. Rather, it expands the number of surfaces on which a brand can be discovered. Those who are technically crawlable, publish original information and answer customer questions precisely create better conditions for this new form of search.
The currently observed retrieval data fits this picture: ChatGPT appears to fan a request out into numerous search paths, check many sources and select only small excerpts for the answer. This increases the value of clearly structured, self-contained sections – but still guarantees no recommendation or citation.
The boundary between optimisation and promise remains important: nobody can guarantee a ChatGPT recommendation. OpenAI itself explicitly says that placement depends on several factors.