AI search optimization means getting a business found and quoted in tools like Google AI Overviews, ChatGPT and Perplexity. You'll also hear it called AI visibility work or LLM optimization. Two narrower jobs sit inside it: AEO, which gets a page's answer quoted, and GEO, which gets a brand described correctly across the sources models draw on. Both sit on top of SEO. A page that isn't indexed has no chance of being quoted.
This page is for agency owners deciding what to deliver and what they can safely promise. SE Ranking estimates about 990 US searches a month for the phrase (modelled estimate, 13 October 2026), so it's a small term, and I wouldn't size the work by it.
What the same question does in practice
Ask the same question in different tools and you won't get the same answer. They name different businesses and cite different sources, and the answers differ in detail too: some are short, some long with links.
Run the same question twice in one tool and a business may be named once and missing the next time, and the sources shown can change too. Starting a fresh session also changes the answer noticeably compared with carrying on an old chat. That's why my panel uses three runs, each in a fresh session (more on that below).
The sources are the part to check, more than the answer. The pages cited weren't always the ones ranking first on Google for the same question. Sometimes an answer leaned on an old or outdated page. Sometimes it cited a directory, a review site or a competitor instead of the business's own site.
Which AI products you can reach
ChatGPT, Google and Perplexity each use their own crawler, so a client can show up in one and not another.
| Product | Examples | What you control |
|---|---|---|
| AI features inside a search engine | Google AI Overviews, AI Mode | Normal Search eligibility: indexed and able to show a snippet. Google documents no additional requirements and no special schema.org markup |
| Chat assistant with search | ChatGPT search | Access for OAI-SearchBot, OpenAI's search crawler. Blocking it keeps a site out of ChatGPT search answers, though the site can still appear as a plain link. GPTBot is a separate setting that concerns training |
| Retrieval-based answer engine | Perplexity | Access for PerplexityBot in robots.txt and, per Perplexity's docs, its published IP ranges. Perplexity says the crawler isn't used for model training |
| Model answering from training alone (my reading, not vendor-documented) | Any assistant with no browsing | Little in the short term |
Sources for this table: OpenAI, Perplexity, Google.
One more OpenAI crawler is worth knowing. ChatGPT-User fetches a page when a person asks for it, and OpenAI says robots.txt rules may not apply to it. Perplexity describes a similar user-triggered fetcher, Perplexity-User, which it says generally ignores robots.txt rules.
My panel runs in ChatGPT, Google AI Overviews and AI Mode, Gemini and Claude. Perplexity is in the table only because it documents its crawler; it isn't in my panel. I have no vendor documentation for how Gemini or Claude crawl or choose sources, so I'm not going to guess.
What the engines are, and why clients keep asking about them, has its own page: AI answer engines.
SEO, AEO and GEO are three jobs with three units of work
A lot of the tactics overlap, but the deliverables and the checks don't, so I'd quote them as three separate lines.
| SEO | AEO | GEO | |
|---|---|---|---|
| Unit of work | A page, for a query | A passage that answers one question | A brand, product or person, and the facts about it across sources |
| What you can check directly | Indexing, rank, clicks | The page is crawlable, the answer is in the raw HTML, the markup matches the visible text | The facts on the client's own properties are correct and consistent |
| What you can only sample | Rank varies by place and person | Whether an engine names the client for a question | How an engine describes the entity across many phrasings |
My rule of thumb for what to sell first: if the client is weak in classic search, start with SEO, because the other two depend on being indexed and understood. If the client ranks but is never named, start with AEO. If the client is named with wrong facts, or confused with someone else, start with GEO. How models retrieve and mention brands is covered in generative engine optimization. Using AI inside your own SEO delivery is a different question, answered in the AI SEO guide.
A top ranking doesn't earn a citation by itself
My working view: Google ranks pages, and the answer engines I check seem to lift one clear sentence, which isn't always from the top-ranked page. A rough test, not a measurement: if you can't find the one-sentence answer on a page in five seconds, a skim-reading tool may struggle too. The page-level version of this problem has its own write-up: why your best page might never get an AI citation.
Google's documentation says AI Overviews and AI Mode may use "query fan-out", issuing several related searches across subtopics. My read: a page can get pulled in for a neighbouring question the client never targeted.
The paper that gave GEO its name, Aggarwal et al., arXiv 2311.09735 (KDD 2024), tested nine content changes on a benchmark of 10,000 queries and also on Perplexity.ai. Adding quotations, statistics and cited sources did best, keyword stuffing gave little to no improvement, and results differed by domain. The headline figure, up to 40%, is a gain on the authors' own visibility metric against their own baseline.
Diagnosing a client who's missing from AI answers
| Symptom | Likely cause | First check |
|---|---|---|
| Absent from AI answers and weak in classic search | Foundation problems | Indexing status, crawl errors, thin or duplicate pages |
| Ranks well, never named | No extractable answer, or no outside corroboration | Is there a plain one-sentence answer near the top? Do third-party pages mention the brand? |
| Named, with wrong facts | Conflicting or outdated information across the web | Put the site, profiles and directory listings side by side and compare the same facts |
| Named in one product, absent in another | Different crawlers, different sources | Robots.txt and firewall rules per crawler; which sources that product tends to cite |
| Competitor named instead | Competitor is better corroborated or described more plainly | Open the sources the product cites for that question and compare them with your pages |
I've seen answers lean on an old page, and I've seen a directory, a review site or a competitor cited instead of the business's own site. The first check is the same each time: open the cited page and read it next to the client's.
Two checks take a few minutes. Swap in the client's domain and a phrase from the answer. The first set is for macOS, Linux or Git Bash on Windows.
# 1. Read the whole robots.txt, then pull out the groups that matter
curl -s https://example.com/robots.txt
curl -s https://example.com/robots.txt | grep -iE -B1 -A3 "^user-agent: (\*|OAI-SearchBot|GPTBot|ChatGPT-User|PerplexityBot|Googlebot)"
# 2. Fetch the page, print the HTTP status, and show a little context around a short phrase
curl -s -A "Mozilla/5.0" -w "\nHTTP status: %{http_code}\n" https://example.com/page | grep -oiE ".{0,60}a short phrase.{0,60}|HTTP status: [0-9]+"
On Windows PowerShell, curl is an alias for a different command (Invoke-WebRequest), so use this form instead:
# 1. robots.txt
(Invoke-WebRequest -UseBasicParsing https://example.com/robots.txt).Content
# 2. Page fetch: status, then context around a short phrase
$r = Invoke-WebRequest -UseBasicParsing -UserAgent "Mozilla/5.0" https://example.com/page
$r.StatusCode
Select-String -InputObject $r.Content -Pattern ".{0,60}a short phrase.{0,60}" -AllMatches | ForEach-Object { $_.Matches.Value }
A crawler with no group of its own follows the User-agent: * group, so read that one first. If the filtered bash command prints nothing, none of those names is listed and the * group, or no rule at all, applies; in PowerShell, scan the printed file for the same names. A missing robots.txt often comes back as an HTML error page, so check the status before you trust the output.
For the page check, use three or four words from the answer. A full sentence often fails on encoded characters (&, ’), wrapped lines or tags inside it. A 403 usually means the server turned the request away, so the result says nothing about the page (PowerShell throws an error instead). Some bot protection answers with a 200 and a challenge page, or a 429 or 503, so read what came back. A 200 with no match means opening view-source: if the words aren't there, the page builds its text in the browser, and a crawler that doesn't run scripts won't see it. A browser-style user agent isn't what a search crawler sends, so this shows what a browser-like visitor gets.
Here's a trap the first check finds. A file that reads User-agent: GPTBot followed by Disallow: / blocks the crawler OpenAI uses for training. It does not block OAI-SearchBot. Because OpenAI treats each crawler as an independent setting, a site can opt out of training and still be eligible for ChatGPT search. The reverse mistake is blocking OAI-SearchBot to "stop AI". That removes the site from ChatGPT search answers.
The smallest AEO fix I use is rewriting a page's first paragraph so it answers the question directly. It takes under an hour for one page.
The order I'd fix things in
An early failure makes later work pointless: with no crawler access, there's no reason to polish copy. Step 7 is the exception, because it's slow and should start early.
- Crawler access. The crawlers you want for search should be allowed in robots.txt and not blocked at the firewall.
- Content in the raw HTML, for crawlers that don't run scripts. The second check above tells you.
- Answer-first sections. Open each important section with the answer, then the support. Start with a handful of pages that match what the client's best buyers ask.
- One version of each fact. Price, location, service list and founding date should read the same everywhere.
- Structured data, if you use it, should match the visible text. Google says no special markup is needed for its AI features, so the aim is consistency.
- Something only the client has. Their own process, data or a method with a name.
- Corroboration: third-party pages that describe the client accurately. Start early, and don't let it hold up steps 1 to 5.
- The neighbouring questions. Pages for the related questions an engine might search for too.
Measuring with a prompt panel
I run a fixed list of 20 to 30 prompts, three runs per prompt, each in a fresh session, on a schedule, across ChatGPT, Google AI Overviews and AI Mode, Gemini and Claude. One run proves nothing, for the reasons in the first section. Write prompts in a buyer's words and mix category, comparison, problem and branded questions. Don't change the list between rounds, or you lose the trend. Run it once before any change and keep that round as the baseline.
Log one row per run: prompt, product, run number, whether the brand is named, whether the domain is cited, whether the facts are right, which competitors appear and which sources are shown.
Read the log prompt by prompt; the total hides the pattern. Some prompts will be stable, with the brand named in every run. Some will be named in one run out of three, and those look like noise until they repeat. Many will be absent in every run, and those are where the diagnosis table applies. After a page change, report a prompt that goes from absent to named in every run and stays there next round.
On the click side, Google's documentation says traffic from its AI features is reported in Search Console under the Web search type, as part of overall traffic. I also use Google Analytics to look for referral traffic from language-model products.
Where tools fit
A professional AI-visibility tool runs the same questions at a larger scale, and I use one alongside the fixed list. The tools measure; they don't move a client up. Treat any score as a sample of one prompt set on one day.
What I won't promise
I decline quotes where the buyer wants guaranteed rankings or traffic. Promising rankings by a date is also one of the four month-one mistakes I see from partners, along with late access, vague scope and slow approvals.
I can't promise a citation, a position or a date. Prompt choice moves any share-of-voice figure, and I've watched one tool name different businesses from one run to the next, so I report mentions and don't guarantee them.
Put three things in the scope: the audit, the page changes and the scheduled AI visibility report. Leave the ranking out. Wording for the clause: "Agency receives the audit, the listed page changes and an AI visibility report on the agreed schedule. Mention and citation outcomes are reported, not guaranteed."
A way to package it
I'd scope it as three steps.
| Step | What the client gets | Service pages |
|---|---|---|
| Audit | AEO: a baseline, the sources engines cite, a competitor comparison and a prioritised fix list. GEO: what each engine says about the company, a factual-error log and a two-list plan. SEO: the classic review | AEO audit, GEO audit, SEO audit; side by side on the audits page |
| Improvement | The page, entity and corroboration changes from the list above | White label SEO agency, white label AEO agency, white label GEO agency |
| Monitoring | An AI visibility report on a schedule: a written report of which questions mention the business, which sources are cited and what to fix. Ongoing access to me on AI search decisions, if you want it | Fractional head of SEO, fractional head of AEO, fractional head of GEO |
On the white label side, my team writes and delivers the audits under your agency's brand, my team checks the work, and I'm your main contact. I never contact your end clients, because that relationship is yours. I've delivered AEO and GEO on a few client sites so far, so treat the packaging as early. Intake agrees scope and collects access and a baseline before the audit begins.
If you're building the delivery side behind this, start at agency operations.
FAQ
What is AI search optimization?
Getting a business found and quoted in tools like ChatGPT, Google AI Overviews and Perplexity. It's SEO plus a few extra jobs.
What is AI search optimization also called?
AI visibility work and LLM optimization are the common ones. AEO and GEO are narrower jobs inside it. I haven't found a formal standard for any of these terms, so ask a vendor what they mean before you compare quotes.
Is it the same as AEO or GEO?
No. It's the umbrella. AEO works on pages so an answer can be quoted, and GEO works on the facts and sources behind a brand. They're separate jobs with separate checks.
Is SEO dead because of AI?
No. Google's own documentation says there are no extra requirements for AI Overviews or AI Mode beyond being indexed and eligible for a snippet. So the SEO basics still decide whether you're in the pool.
Do I need special schema markup to appear in Google AI Overviews?
Google's documentation says no special schema.org structured data is needed for its AI features. Keep markup accurate, because it should match the visible text.
What tools do I need to measure AI search visibility?
A fixed prompt list you can run by hand is enough to start. Tracking tools run the same method at larger scale, and Search Console and analytics cover the traffic side. Tools sample answers, so none of them proves a position.
How long does AI search optimization take?
I can't give a timeline. A single page edit can take under an hour. When a product shows the change depends on how it crawls, and corroboration depends on other sites, so it's the part you control least.
Keep reading
- AI answer engines: the five-minute check, and a paragraph you can paste to a client.
- Generative engine optimization: how models retrieve and mention brands, and what the research shows.
- Why your best page might never get an AI citation: a single-page failure and the edit that fixes it.
- AI SEO: what AI means for your SEO work and tools.
- Audits: SEO, AEO and GEO side by side
Sources: OpenAI crawler documentation · Perplexity crawler documentation · Google: AI features and your website · Aggarwal et al., GEO: Generative Engine Optimization, arXiv 2311.09735 · SE Ranking (US).
