v(i)b(e)c(o)d(e)4u.com
Назад на форум vbcd4u.com
🔥 Нейросети2026-11-17T09:23:00.000Z▲ 0 upvotes

AEO without myths: what really affects citations in Perplexity and ChatGPT

@murdasoftРедакция vbcd4u · MurdaSoft
Открыть все обсуждения →

First, about what no one promises

There is no way to be guaranteed to get answered by Perplexity, ChatGPT or AI Overviews. Models select sources according to their own rules, which are not published and change. Anyone selling “guaranteed ChatGPT quotes” is selling hope.

What can be controlled: so that the search robots of these systems can pick up the page, so that it has a clear answer to a specific question and so that this answer is easy to retrieve. This overlaps with regular SEO more than people say.

Step 1. Do not close yourself off from the necessary robots

Each company has several agents with different tasks. Blocking “all AI bots” with one switch often closes the search, which just gives quotes.

  • OpenAI: OAI-SearchBot - index for searching in ChatGPT, GPTBot - collection of data for training, ChatGPT-User - page loading at user request.
  • Perplexity: PerplexityBot for index and Perplexity-User for user queries.
  • Anthropic: ClaudeBot, Claude-SearchBot, Claude-User with similar division.
  • Google: AI Overviews are built on the regular Googlebot index. The Google-Extended token controls the use of content for Gemini, but does not affect placement in Search.

Example: search is allowed, training is prohibited.

# robots.txt
User-agent: OAI-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

User-agent: GPTBot
Disallow: /

User-agent: *
Allow: /
Sitemap: https://example.com/sitemap.xml

Then check the layer in front of the site. Cloudflare has managed rules for AI crawlers, and if they were once enabled “just in case,” robots.txt no longer solves anything—requests are cut off by the WAF.

Step 2. Check that the robot sees the text

Many crawlers, judging by the logs of site owners, take HTML and do not execute JavaScript. If the response only appears after hydration, the robot sees an empty shell.

curl -s -A "OAI-SearchBot/1.0" https://example.com/ru/forum/some-thread \
  | sed -e 's/<[^>]*>//g' | grep -c "ключевая фраза из ответа"

Zero - the text is rendered on the client. In Next.js, this means that the content should be served from the server component, and not loaded via useEffect.

At the same time, look at the status and response time: a page that responds in 4 seconds or returns 403 to unknown agents loses to fast and open ones.

Step 3. Write so that the answer can be cut out

Search engines cite fragments. It is easier to take a fragment if it is self-sufficient.

  • Formulate the section title as a question or as an exact statement: “Why Docker bypasses UFW”, and not “Nuances”.
  • The first one or two sentences under the title are a direct answer. Details, reservations and code are below.
  • Specifics instead of general words: version, command, default value, limitation. “Works starting from nginx 1.25” is quoted, “in modern versions” is not.
  • Indicate the update date and author. For technical matters, freshness is important for both people and machines.
  • Comparison tables and lists of steps survive retrieval well if each line has a complete thought.

This is not an AI trick: such pages are better read by people and rank better in organic search.

Step 4. Structured data - to the point

JSON-LD helps the machine understand the page type, author, dates, and hierarchy. There is no public evidence that markup itself increases the chance of citations in LLM responses. Make it correct and relevant to the visible content, and not for show: Article or DiscussionForumPosting with real dates and author, BreadcrumbList for navigation.

About llms.txt

llms.txt - a proposal (Answer.AI, 2024) to place a Markdown file with a brief description and links to key pages in the root of the site. This is not a standard, and major search engines have not stated that they take it into account when selecting sources; Google representatives said they don't use it. The file is cheap to make and can be useful to agents to whom the user has clearly given your site. But there is no reason to expect an increase in citations from him.

How to measure

Without measurement, AEO becomes a belief. Three data sources:

  1. 1.Server logs. How many requests from search agents and what pages they retrieve.
grep -E "OAI-SearchBot|PerplexityBot|Claude-SearchBot" /var/log/nginx/access.log \
  | awk '{print $7}' | sort | uniq -c | sort -rn | head -20

User-Agent is easy to fake. OpenAI and Perplexity publish IP ranges for their robots - check them before jumping to conclusions.

  1. 2.Transitions. ChatGPT adds utm_source=chatgpt.com to links, transitions from Perplexity are visible by referrer perplexity.ai. Create a separate segment for them in analytics.
  1. 3.Manual Sampling Once a month, ask the systems 20 questions to which you have the best answer and make a note of who gets cited. It's rough, but it shows the dynamics and competitors.

Common mistakes

  • Close GPTBot and think that you closed the ChatGPT search, or vice versa.
  • Generate hundreds of “question-short answer” pages under AEO. Such pages rank poorly both in classic search and in systems that rely on it.
  • Hide the response behind a modal subscription window or cookie banner that overlaps the content in HTML.
  • Expect the effect in a week. Indexes are not updated instantly, and single observations do not prove anything.

When not to do this

If the site does not have indexing in regular search, you need to start there: many response systems rely on search indexes. If the product solves problems that are not asked about in text (internal tools, closed B2B systems), investments in AEO will not pay off.

Обсуждение (0)

Ещё на vbcd4u.com