Manual
Product site My Page

AI Recommendation (LLMO)

Last updated · Published · Sorabun Editorial Team

Get recommended when people ask ChatGPT, Gemini, Claude, Perplexity or Google's AI Mode. Define customer questions (CEP), find the AI's selection criteria (KBF), and deliver your reasons to believe (RTB) to articles and outside media. Also shows mention and citation rates per competitor

Ranking high in search results is not enough if ChatGPT, Gemini, Claude, Perplexity and the like leave your name out when someone asks "which one do you recommend?" "Sorabun > Leave It to AI > AI Recommendation (LLMO)" walks you through four steps to get chosen by AI on one screen (Pro, administrators only).

StepWhat you do
1. Design the questions (CEP)Decide what customers ask AI (their situation, problem and conditions)
2. Identify the criteria (KBF)Actually ask AI and find out what it uses to compare candidates
3. Organize the evidence (RTB)For each AI criterion, check whether you show facts and data that make you the right choice
4. Place it where AI looksDeliver missing evidence to your own site (articles) and to outside media that AI reads

AI builds answers from two sources: pre-training (learned knowledge and mentions on the web) and retrieval (RAG) (outside information, comparison media and official sites it searches while answering). This screen shows the steps to get your evidence into both.

1. Design the questions

  • Your name: when this name (or one of its aliases) appears in an AI answer, it counts as "recommended". The name is never included when asking AI (that would skew the answer and make the measurement meaningless). Your site's domain (example.co.jp) and its stem (Example) are also looked for as your name (see "How names are found" below)
  • Customer questions: one per line, up to 30. Write them the way people talk to AI, not as search keywords. Include plenty of questions where AI lists candidates, like "which one do you recommend?"
  • "Have AI write questions": creates about 15 questions from your Site Diagnosis and Rank Tracking keywords. Edit them closer to your real customers' words before saving

2. Ask AI and find the selection criteria

Press "Ask AI and measure" to ask each question to every AI you have an API key for (see the table below), three times each (three times since 2.133.0; before that, once each). From each answer it reads:

  • The companies and services named as recommendations (in order), and where you appear among them
  • The criteria used to compare candidates (fees, support, track record and so on)
  • The sites used as sources. When your site is cited, which article (URL) was cited

You then see the share of answers that named you (overall and per AI), how many of the three answers named you for each question, which of your articles were cited, and each question's answer (you can open the full text). Every measurement adds to the overall history and to a per-question history (the latest 12 measurements).

You also see the following (since 2.136.0). All of them are counted from the answers already collected; Sorabun asks AI nothing extra for them.

  • Mention and citation rates per competitor: a table, next to your own row, of the share of answers that named each competitor and the share whose sources included the competitor's site (subdomains too). Competitors come from "Competitors" in Site Diagnosis and from the names AI recommended. Citations are counted only for competitors whose domain is known (write a domain or URL, not just a name, under "Competitors" in Site Diagnosis to count them)
  • Questions with your name (counted separately): questions that contain your name or domain ("What do people say about X?") will of course name you. Mixing them into the total would inflate the share, so they are left out of the overall and per-AI shares and counted separately. They are marked "Contains your name (counted separately)" in the list of questions
  • Searches AI ran behind the scenes: the search terms AI ran while building its answer (fan-out), and in how many answers. Shown only when the AI's response includes them (the Gemini, ChatGPT, Claude and Perplexity APIs, and ChatGPT's on-screen answer; Gemini's on-screen answer and AI Mode do not include them). Check whether your pages come up when you search these terms
Why three times, and why no ranking

An AI's answer changes every time, even for the same question (the names it lists change). After asking only once, you cannot tell "named" or "not named" from chance. So Sorabun asks the same question three times and counts how many of the three answers named you, and adds a range to the overall share (the range the true share is likely to fall in, statistically). While the range is wide, do not read the difference from last time as "up" or "down."

For the same reason, Sorabun does not show your "rank" inside AI as a single number (such as an average position). It shows only, for each answer that listed you, where you were in that answer.

Which AIs are asked

Sorabun asks only the AIs you have an API key for. Each gets the question as is, the way an ordinary user would ask (it never adds "search the web" or "cite your sources").

AIKeyHow it is askedWhat counts as a sourceEstimated cost per question
Geminigemini_api_keyThe Gemini model chosen under "Models", with Google SearchThe search results the answer was grounded onDepends on the model
ChatGPTopenai_api_keyThe GPT model chosen under "Models", with web searchThe citations attached to the answerDepends on the model
Claudeanthropic_api_keyClaude Sonnet 5.5, with web search (up to 3 searches per answer)The citations attached to the answerAbout $0.08
Perplexityperplexity_api_keyPerplexity's Sonar, with web searchThe search results listed as the answer's sourcesAbout $0.006
ChatGPT (on screen)dataforseo_login + dataforseo_password and aiscrape_chatgptDataForSEO fetches the answer shown on chatgpt.comThe sources the answer drew onAbout $0.0012
Gemini (on screen)The same credentials and aiscrape_geminiDataForSEO fetches the answer shown in GeminiThe sources the answer drew onAbout $0.0012
Google AI ModeThe same credentials and aiscrape_aimodeDataForSEO fetches the answer in Google's AI ModeThe answer's references and the links in the answerAbout $0.0012
  • Claude is asked with Claude Sonnet 5.5, separate from the model that writes your articles (Claude under "Settings > Models"). Anthropic makes the Sonnet line the default on claude.ai for the Free and Pro plans, so its answers are close to what ordinary users see. Choosing a higher model for writing does not raise the cost of measuring
  • Perplexity is asked through Perplexity's Agent API, naming Perplexity's own search model, Sonar (the old Sonar API reached end of support in September 2026)
  • The estimates come from the prices on each provider's pricing page, checked in November 2026 (Claude: 2 searches, about 20,000 input tokens and 2,000 output tokens; Perplexity: 1 search, about 7,000 input tokens and 800 output tokens). They change with the number of searches and the length of the answer
  • When an AI declines to answer (such as a Claude safety decline) or answers without searching, the answer counts as "answered, no sources". Only times when no response came back from the AI count as "could not ask"

On-screen AI answers (DataForSEO)

An answer from a provider's API can differ a little from what ordinary users see on chatgpt.com or in Google Search (the screens add their own search and display). With DataForSEO credentials you can also measure with the answer and sources actually shown on screen (since 2.136.0).

  • How to use it: under "Settings", in "Search Data Source", choose ChatGPT on screen, Gemini on screen or Google's AI Mode in "AI answers on screen (DataForSEO, optional)". All are off by default. It uses the same credentials as search volume, dataforseo_login and dataforseo_password
  • How it asks: it fetches answers for Japan in Japanese. The question is sent as is (your name is never added)
  • Timing: questions are queued first and the answers are read later (DataForSEO's standard queue, up to about 45 minutes per answer). The second and third rounds are queued 30 minutes after the previous round (collecting answers from the same moment would understate how much they vary). So it takes about one to two hours until all answers are in. Even if you close the screen, it continues every hour
  • Cost: about $0.0012 per answer (DataForSEO pricing, standard queue, November 2026). 15 questions asked 3 times cost about $0.05 per engine. The screen shows the estimate for the measurement and how much is left for today
  • Daily limit: aiscrape_daily_usd (default $1). Questions beyond the cap are not queued and show as "Could not ask" (daily limit). Even after a failure where it is unclear whether you were charged (such as a lost connection), the same question is never queued again (so you are never charged twice)
  • Answers DataForSEO could not produce, and answers that have not arrived after 12 hours, count as "could not ask" (they are left out of the denominator)
  • DataForSEO's documentation examples did not confirm that Japan is available for Gemini's on-screen answer. If it is not, those answers show as "could not ask" with the reason

How names are found

These rules apply when looking for your name (and competitors' names) in answers (since 2.136.0).

  • Names inside URLs and email addresses do not count (https://example.jp/sorabun or [email protected] is not a mention). Source markers in the answer ([1] or [example.jp]) do not count as names either
  • Full-width and half-width characters (letters, digits and katakana) and upper and lower case are treated the same (normalized with Unicode NFKC)
  • Names made of letters and digits are matched at word boundaries. "Acme" does not match "Acmeride", "MyAcme" or "acme-pro"
  • Katakana-only names of up to 3 characters do not match inside a longer katakana word ("ソラ" does not match "ソラリス"). Names in kanji or hiragana still match anywhere in the text, as before (Japanese has no spaces between words)
  • Your site's domain and its stem are also your name (for acme.co.jp, both acme.co.jp and "Acme"). Stems of 3 characters or fewer, and common English words such as shop or news, are not used. A site running on a subdomain does not use a stem
  • When checking sources, subdomains count as your site (for example.jp, a citation of blog.example.jp is a citation of your site). AI citation check works the same way
About cost

One measurement makes "questions × AIs × 3" queries, the same number of reads, plus one final analysis (about 181 calls for 15 questions × 2 AIs, or about 361 calls with 4 AIs). Costs follow your API key pricing, and queries with search cost more than usual. For the questions to Claude and Perplexity, the screen shows an estimated amount (for 15 questions, about $3.60 for Claude and about $0.27 for Perplexity). "Re-measure once a week with the same questions" is off by default.

On the Cloud edition, Gemini and ChatGPT are asked through the Sorabun server and use that much more of your allowance. Claude and Perplexity are asked only when you enter your own API keys on the Settings screen (the Sorabun server cannot call them, so your site asks each provider directly). You pay for those questions from your own account.

The DataForSEO part never goes over the daily limit (default $1). One measurement of 15 questions × 3 times costs about $0.05 per engine.

It asks one query at a time, so with many questions it takes tens of minutes. Even if you close the screen, it continues every hour. If it stops partway, what has been asked is kept and it resumes from there.

Results measured up to 2.132.0 are read as "N out of 1" (the screen shows a notice suggesting you measure again).

3. Evidence for each selection criterion

The collected criteria are grouped into 4 to 8 "AI selection criteria", and your evidence is checked for each. Evidence comes from the Evidence Lab, first-party data, Site Diagnosis and published articles.

ResultMeaning
EvidenceShown with facts, numbers or third-party reviews
Weak evidenceMentioned, but without numbers or sources
No evidenceYour site does not show why you would be chosen on this criterion

"Evidence to publish" tells you what to make public. Making up evidence is never recommended. Register real numbers, cases and third-party reviews in the Evidence Lab with "Register evidence", then turn them into an article with "Write an evidence article".

4. Place it where AI looks

  • Media AI cites: sites AI reads while answering. Aim to get your evidence onto them through listings in comparison articles, interviews, guest posts or reviews
  • Others AI recommended: how many times each was named. Looking at what evidence they publish for the criteria above shows what you are missing
  • Your own site: set evidence articles to the "LLMO" or "Hybrid" optimization type. Keep structured data (JSON-LD) on so the author's job title, the reviewer, and the sources are told to search engines accurately (SEO Plugin Integration and Structured Data). Google says no special structured data or llms.txt is needed to appear in its AI features. Whether AI actually reads your site, and whether people arrive from AI, can be checked on Visits from AI

Used together with AI citation check (whether your site is among the sources for each keyword), you can see both "are you read as a source" and "are you named as a recommendation".

How names are found, mention and citation rates per competitor, counting questions with your name separately, and fetching on-screen answers through DataForSEO are based on the design of the open-source SEO tool OpenSEO (MIT License, (c) 2026 Ben Senescu), rebuilt to run inside WordPress.