Research guide · GEO
GEO Generative Engine OptimizationIn AI answers,
does your brand appear?
GEO (Generative Engine Optimization) is content improvement focused on brand presentation and source citations in AI answers, covering website information, verifiable sources and consistency with external records.
Explore retrieval principles, MCAC observation methods and citation checks, keeping brand mentions, website citations and factual accuracy separate.
01
GEO reading guide
Start with retrieval and real observations, then examine traffic and citation sources.
- Growth
From vectors to citations: semantic retrieval behind GEO
Understand how semantic retrieval finds candidates and how retrieval differs from citation.
- Research
How AI search finds local businesses
Use observations of Taipei beef noodle restaurants to examine mentions, citations and accuracy separately.
- Growth
From Referral to AI Assistant: how GA4 classifies AI traffic
Identify AI-referred visits and distinguish them from answer exposure that analytics cannot observe.
- Research
What happens when AI starts citing AI-generated articles?
Trace cited sources to their evidence and examine how AI-generated content enters knowledge sharing.
Related terms: GEO · Embedding · RAG · Simpson’s paradox.
Further reading: How AI compares brands: retrieval and ranking (Traditional Chinese).
02
Four MCAC checks
MCAC is LS-Labs’ diagnostic framework, not a universal GEO score or a search platform ranking standard. It records mentions, citations, accuracy and conversion-readiness separately. These measures cannot be added into one brand score: each identifies a different issue. See the definition of MCAC and the methods and sample in How AI search finds local businesses.
- 01
Brand mention Mention
Does the answer mention you?
First identify the brand being discussed, then check whether it appears in a recommendation, comparison or criticism. Record prompts that already name the brand separately.
- 02
Source citation Citation
Which page does it link to?
Record the links, footnotes or source list and check whether each source supports the claim. A brand mention is not necessarily a citation of its website, and visible citations do not reveal every source the model used.
- 03
Information accuracy Accuracy
Are the facts about you correct?
Check facts such as addresses, phone numbers and prices against valid records. Keep correct, incorrect, missing and unverifiable information separate.
- 04
Conversion-readiness Conversion-readiness
Does the answer offer a next step?
Check whether the answer provides a phone number, URL, address or another next step. This does not establish that someone contacted the business, booked or purchased.
03
Checking a restaurant answer
Apply all four MCAC checks to the same restaurant answer. Switch between the scenarios to see whether Restaurant A is mentioned, which sources are cited, whether the facts are correct, and whether a next step is provided.
User request“Find a restaurant and book a table for two tonight at 19:00.”
Your name appears in the answer
Consider Restaurant A, a possible choice for dinner for two. Check tonight’s availability before booking.
The answer includes the restaurant’s name
What to checkCount the appearance of the name as a brand mention. Then inspect the context, such as a recommendation or comparison, and the links supplied.
Sources cited in the answer
Restaurant A offers online reservations.Source
What to checkOpen the cited sources to check whether the information comes from an official website, media outlet or directory. The brand mentioned may not own the website cited.
The source must support the answer
Missing evidence: live availability
What to checkOpening hours do not establish tonight’s availability. That requires checking the booking system; asserting availability without it is unsupported.
Does the answer provide a next step?
Restaurant A offers online reservations and accepts enquiries by phone.
The answer provides a booking page and phone number
What to checkCheck whether the answer provides a phone number, URL, address or another next step. An available booking link does not establish that a booking occurred.
How are sources retrieved? Explore the diagram ↓
04
How vector similarity works
Vector retrieval can use similarity between text vectors to find candidates. This two-dimensional teaching model uses booking information and restaurant stories to illustrate direction and angle. Its angles and ranks are illustrative, not measured Google or AI platform rankings.
Your page says
“After ten years training in Tokyo, our chef returned to Taiwan and built this restaurant with reclaimed timber, hoping to give every guest a taste of home.”
“Our chef trained in Tokyo for ten years. The restaurant seats twenty. Call us if you would like to visit.”
“Dinner runs from 17:30 to 21:30 in 30-minute slots. Parties of two or more can book on our website, which shows the tables available that evening.”
The directions differ: similarity is lower in this example.
Ranked 4th in this model, outside the top 2 candidates.
- A Booking platform listing0.990
- B Food blogger’s review0.866
- C Magazine chef interview0.574
- d Your page0.469
This model only compares preset directions. Real retrieval also depends on the query, index and ranking method; retrieved content is not necessarily cited in the answer.
cos θ = (q · d) / (‖q‖ ‖d‖); 1 at 0°, and 0 at 90°
- 01
Represent text as vectors
An embedding model represents queries and passages as lists of numbers, illustrated here as arrows. Semantically similar text tends to have similar vector directions.
- 02
Compare angles and retrieve candidates
This model sorts by cos θ and retrieves the top two passages. Real similarity methods and candidate counts vary by implementation.
- 03
Generate an answer from candidates
Candidates still undergo selection and answer generation. Clear conditions, figures and sources help readers verify content, but do not guarantee citation.
05
Common questions
Who introduced GEO?
The academic term comes from Aggarwal and colleagues’ paper, GEO: Generative Engine Optimization. Its preprint appeared in November 2023 and it was later accepted at KDD 2024, introducing a content optimization and visibility evaluation framework.
How does GEO differ from SEO?
SEO examines search rankings and clicks; GEO examines brand mentions, source citations and factual accuracy in AI answers. According to Google Search Central , pages must be indexed and eligible for snippets to appear in Google’s AI features. SEO fundamentals still apply, while mentions and accuracy require their own checks.
Does a brand mention mean its website was cited?
Inspect the answer’s links. A brand can be mentioned while the cited sources are its website, media, directories or other sites.
What is MCAC, and how does it relate to GEO?
MCAC is LS-Labs’ diagnostic framework for recording mentions, citations, accuracy and conversion-readiness separately. It is a diagnostic approach, not a universal GEO scoring standard.
How can I evaluate a content change?
Keep prompts and evaluation rules fixed. Record the platform, model, date, full answer and sources, and repeat observations. Results can vary across questions, platforms and time; one mention does not prove that an edit worked.
Can the four MCAC measures become one score?
No. They measure different things and must be compared separately. A combined brand score hides whether the issue is a mention, citation, accuracy or an available next step.
Can a business appear without its own website?
Yes. In LS-Labs’ observational sample of six Taipei beef noodle restaurants, AI also mentioned businesses without websites. Visible sources included business profiles, food articles and reviews. An official website is a first-party source the business can maintain; the sample does not prove it improves answer accuracy.
06
How the four research topics connect
GEO examines mentions and citations in organic answers, TSO examines which tool an agent selects, AX examines whether a task can be completed, and AI Ads examines how paid placements below an answer are matched and converted. These four topics address adjacent problems, so you can read across them as needed.
