AGENTICOPTIMIZATIONGEO / AEO / LLMO
Is AI telling buyersthe right things about you?
So that when someone asks an AI, your company is named and your product stays on the shortlist. We make your company information and product data readable by AI, measure your mentions in AI answers every month, and keep correcting it. We put the same questions to six AIs, grade every answer against your official facts, and tell you exactly what to fix.
ChatGPT
Gemini
Claude
Perplexity
Google AI Overviews- AI Mode
Six AIs, measuredChatGPT, Gemini, Claude, Perplexity, Google AI Overviews, AI Mode
Is the ORBIT O-12 waterproof? We'd use it on a dusty shop floor.
Sourcesrugged-guide.example · orbit.example
Answer checkClaims found
CheckedDust- and waterproofMatch
OfficialDust- and waterproof
Rated IP68Mismatch
OfficialIP67orbit.example/o-12/spec
Passed a 1.5 m drop testMatch
Official1.5 m drop test
States a protection rating that differs from the official spec
Fix
List the protection rating as its own row in the spec table
The same question, put to 6 AIs
Faults found in 6 of 6 answers
- ChatGPT: Yes. The O-12 is a rugged tablet that is dust- and waterproof, with an IP68 rating. It has also passed a 1.5 m drop test. (Wrong)
- Gemini: The O-12 is rated IP67 for dust and water, and it is certified to MIL-STD-810H. (Wrong) Red: wrong spec. Claims a certification the official spec does not list. Fix: State which tests the O-12 passed, and that it has no MIL-STD certification.
- Claude: For a dusty floor, pick a rugged tablet with a high IP rating. Here are a few to compare. (Not mentioned) Grey: not mentioned. Leaves the brand out of an answer its product fits. Fix: Answer 'is it waterproof?' in one plain sentence on the product page.
- Perplexity: Yes, the O-12 is IP67. It weighs 1.4 kg, so it is easy to carry around the floor. (Wrong) Red: wrong spec. States a weight that differs from the official spec. Fix: Give the weight with and without accessories in the spec table.
- Google AI Overviews: The ORBIT O-12 is an IP67 rugged tablet with Wi-Fi 5. (Wrong) Red: out of date. Repeats a spec from an older version of the product. Fix: Date the spec page and say which batches changed.
- AI Mode: Yes, it is IP67, but its screen is hard to read in direct sunlight. (Wrong) Red: wrong spec. Contradicts the official display spec. Fix: Say in a sentence what the 1,000-nit display means outdoors.
Illustration only. ZAG wrote these answers to show how a check works; they are not real AI output, and ORBIT is a fictional laptop brand.
Example: fictional brand ORBITPrompt design
Prompts written on a hunch
can't give you a reliable measure.
ZAG designs personas and prompts from five sets of real data, then shows how AI recommends you against objective criteria.
What the animation shows
- Design personas and prompts from five real data sources.
- Cover every persona × intent, so nothing is missed or skewed.
- No padding. Stop at the saturation point, where no new issues appear.
- Feed each month's measurements into the next round of prompt review.
Method
Six steps:
measure, fix, measure again.
- 1
Build the question set
From Google search demand, AI search volume (estimated), how people actually phrase questions, official facts and competitors. Each question tagged with persona, buying stage, intent and stakes.
- 2
Ask six AIs
ChatGPT, Gemini, Claude, Perplexity, Google AI Overviews and AI Mode, keeping every answer and citation.
- 3
Grade for accuracy
Was it right, not only were you mentioned: wrong price, wrong spec, a rival recommended instead. The grading itself is audited.
- 4
Map demand and rivals
Where the demand is, and who wins it against each rival.
- 5
Pinpoint the fixes
The few fixes that matter most: what to change, on which page, and why.
- 6
Measure again
The same questions, re-measured to prove the fix worked.
Citations, the rivals AIs name and events found while tracking feed back into the question set.
Grading
Accuracy,
not mention counts.
Being named doesn't help if the price is wrong, the spec is wrong or a rival gets recommended.
- written rules
- 15+
- Answers graded against an official answer key
- grading audit
- 2-layer
- A blind re-grade, plus a second-model review of every red
- answers and citations kept
- All
- Append-only, never overwritten
Buyers are moving their product search
from search engines to AI.
When buyers search with AI, only companies whose product data is in a form an AI can read make the shortlist.
Source: our audits of websites, mostly Japanese manufacturers (search keywords February 2025 to July 2026; mentions in AI answers September 2025 to July 2026). See the method and the distributions →What you receive
Not just numbers. The fixes too.
- 01
The question set
Tagged by persona, buying stage, intent and stakes.
- 02
Graded answers
Six AIs' answers, their citations and the reason for each verdict.
- 03
A map of demand and rivals
Where the demand is, and which rivals win it.
- 04
Fix proposals
Which page, what to change, and why.
- 05
Re-measurement
The same questions, before and after the fix.
How to start
Start with the free audit.
- 01
Free audit
Enter a URL and get a quick report, on the spot, scoring what public information shows.
- 02
Phase 0
A check on your real data, about 2–3 weeks after you provide the materials.
- 03
Ongoing measurement
The same questions, measured again to prove what worked.
Pricing is quoted individually, based on the product categories, data scope and how often it runs.
Questions
- GEO, AEO, LLMO: what's the difference?
- They are names for the same goal: being represented correctly in AI answers. We don't split them; we measure six AIs' answers and fix what's wrong.
- Which AIs do you measure?
- Six: ChatGPT, Gemini, Claude, Perplexity, Google AI Overviews and AI Mode.
- Isn't mention share enough?
- Being named doesn't help if the price is wrong, the spec is wrong or a rival gets recommended. We grade correctness against your official facts.
- How is it priced?
- By individual quote. The free audit is the place to start.
How do your site and products look
to your customers' AI?
Send us a URL and we show you an instant report on the spot, scoring what can be verified from public information. A full report, including a comparison with competitors, can be requested free of charge.
Submitting sends the domain you entered, and the name that site publishes for itself, to our measurement vendor to count mentions in AI answers. Nothing about you is sent. By submitting, you confirm you may have this site assessed and that the choice of site is yours (terms, terms-v1.2 (2026-10-04)).
