How to Optimize for Perplexity, Gemini and Claude
To optimize for Perplexity, Gemini and Claude, do the same work that gets you named in
ChatGPT, then check one thing per assistant: whether the specific crawler it sends is
allowed to read your site. Each company runs two or three separately named bots with
separate jobs, and your robots.txt file treats every one of them as a different visitor. A
rule written to keep your content out of AI training can quietly keep your business out of
AI answers, and most owners never look.
Key Takeaways
- Optimizing for Perplexity, Gemini and Claude is mostly the same job as ChatGPT. They
differ in how they reach your site and whether the user sees your name, not in what they
reward. - Each company runs several crawlers with different jobs. OpenAI documents four,
Anthropic three, Perplexity two. Blocking the wrong one removes you from answers. - Google-Extended does not control Google Search. Google states it governs training and
grounding for Gemini models, and is not used as a ranking signal in Google Search. - None of the four publish how they score sources. Any article quoting you a citation
percentage or a freshness weight is quoting a guess.
If you have asked ChatGPT who to call for your trade in your own town and seen your name come back, you are probably wondering whether Perplexity, Gemini and Claude would say the same, and whether you need to optimize for Perplexity as a separate job. This guide goes assistant by assistant: the crawler each one sends, the file that can quietly lock you out, and a check you can run this week. I have spent more than 15 years in search, and robots.txt is the first file I open on a new site.
Who is actually using the other assistants?
In a Pew Research Center survey fielded in February 2026, about half of U.S. adults said
they use AI chatbots. ChatGPT led at 44% of adults. Gemini reached 24%, Copilot 17%, Meta
AI 14%, Grok 8% and Claude 6%. Those figures come from
Pew's 2026 report
on American AI use.
In our AI gap report analysis of 17 local businesses across 1,352 AI answers, we found when at least one engine named a business, every engine agreed only 9% of the time, 10 combinations out of 110.
Read that as an owner rather than a marketer. ChatGPT is the biggest room, and it is not
the only room. Roughly one in four adults is asking Gemini. When a homeowner with a failing
water heater asks any of them who to call, the answer comes back as two or three names, not
a page of ten blue links. Being absent from one assistant is not a rounding error when the
shortlist is that short. Our guide to
AI visibility for local business
covers how that shortlist gets built.
The assumption worth retiring is the one most owners make after a good result: they type
their trade and their city into ChatGPT, see their own name, and treat the job as finished.
On the surface that makes sense, because ChatGPT is the biggest room and seeing your own name looks like proof. But that test covers one assistant on one day. It says nothing about whether Perplexity can
reach the site at all.
We saw how thin that proof is when we added up four of our AI Recommendations Gap Reports from September 2026. Each asked one question, which company would you recommend for this service in this city, across every service, city and engine. Of 285 queries, 277 returned an answer, and 100 of those named the business: 36.1%. Three of the four businesses had the most mentions in their own market and were still missing from between 25% and 57% of answers. An engine that returned nothing was dropped from the count, so every miss is a real answer that came back without the business in it. Leading your field and being reliably recommended turned out to be separate conditions.
If you want a second opinion on where your business stands, that is what a conversation
with us usually starts with.
How does each assistant actually find your website?
A crawler is an automated visitor that reads your pages. These four companies send crawlers
to do three different jobs, and the difference matters because you can allow one and refuse
another.
The first is training, collecting text that may be used to build future versions of the
model. The second is search indexing, building the catalog the assistant searches when
someone asks a question. The third is a live fetch, when a person asks something right
now and the assistant reads a page to answer. Blocking one job does not block the others.
Here is what each company documents about itself, as of September 2026.
| Assistant | Crawlers it sends | How it reaches your page | What controls access | How your name shows up |
|---|---|---|---|---|
| ChatGPT | GPTBot, OAI-SearchBot, ChatGPT-User, OAI-AdsBot | OAI-SearchBot surfaces sites in ChatGPT's search features; ChatGPT-User visits when a person asks | robots.txt, per bot. Opting out of OAI-SearchBot removes a site from ChatGPT search answers | Linked sources in the answer |
| Perplexity | PerplexityBot, Perplexity-User | PerplexityBot surfaces and links sites in Perplexity's search results; Perplexity-User visits on a user's request | robots.txt for PerplexityBot; Perplexity-User generally ignores it because a person triggered the visit | A link to the page included in the response |
| Gemini and Google AI surfaces | Googlebot, Google-Extended | Google's search index. A page must be indexed and eligible to show with a snippet | robots.txt for Googlebot, plus nosnippet, data-nosnippet, max-snippet and noindex. Google-Extended covers training and grounding only | Supporting links beside the response |
| Claude | ClaudeBot, Claude-SearchBot, Claude-User | Claude decides per question whether to search. Claude-SearchBot works on search quality; Claude-User visits when a person asks | robots.txt, honored per bot | Citations carrying the source URL, page title and a short quoted passage |
Run your eye down the fourth column. That is the only one you control directly this week,
and it is where nearly every self-inflicted visibility problem lives. A developer who added
a broad AI block in 2024, or a security plugin that refuses unfamiliar user agents, can take
a business out of three assistants without anyone noticing. Nothing breaks. No error
appears. The phone is quieter than it was, and the cause is a text file almost nobody opens.
One caution before you read further on this topic elsewhere. None of the four publish how
they score or order the sources they pick, so the table is the documented part.
How to optimize for Perplexity?
To optimize for Perplexity, make sure PerplexityBot can read your site, because that is the crawler that puts websites into Perplexity's search results. The proof is the most useful sentence Perplexity has published
about itself. PerplexityBot, its documentation says, is "designed to surface and link
websites in search results on Perplexity. It is not used to crawl content for AI foundation
models." Read that twice if you have ever been advised to block AI bots. A rule aimed at
keeping your content out of AI training does not remove you from Perplexity. A rule aimed at
PerplexityBot does, and it is the rule most blanket blocks accidentally write.
Perplexity's crawler documentation names both
agents.
The second agent is Perplexity-User, which supports actions inside Perplexity. When somebody
asks a question, it may visit a page to help answer it. Because a human triggered that
visit, the documentation says the fetcher generally ignores robots.txt rules, and Perplexity
says it will include a link to the page in its response. A user-initiated visit can surface
your name even where an indexing rule would not.
Perplexity also publishes the IP address ranges both agents use. That matters more than it
sounds. Anything can claim to be PerplexityBot in a server log, and the published ranges are
how you tell a real visit from a bot wearing someone else's name. If your host has ever
blocked "suspicious" traffic in bulk, this is where you check what got caught.
Access questions like this are the fastest part of our free audit to answer.
What does Gemini do differently?
Gemini is the cheapest of the four to cover, because you are probably already doing the
work. Google's guidance on its
AI features states that
"to be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be
indexed and eligible to be shown in Google Search with a snippet," and that there are "no
additional requirements … nor other special optimizations necessary." Ordinary indexing
plus snippet eligibility is the entry ticket, though Google adds that meeting every
requirement still does not guarantee it will serve your content.
The confusion worth clearing up is Google-Extended. If you want your content kept away from Google's AI, it looks like the switch to flip, and on the surface that makes sense, because the name reads that way. It is not. Google's
crawler documentation
describes it as a control for whether crawled content may be used for training future Gemini
models and for grounding, then states that "Google-Extended does not impact a site's
inclusion in Google Search nor is it used as a ranking signal in Google Search." Turning it
on is a decision about model training, and treating it as a way to manage how you appear in
Google's answer surfaces leads owners to fix the wrong file.
So what people call Gemini SEO is, for a local business, mostly the Google work already in
progress: get indexed, stay snippet-eligible, keep the facts about your business correct.
Worth a short call if you are unsure which of those three you are missing.
What does Claude do differently?
Claude has the most distinctive behavior of the four, and it has nothing to do with content
format. Anthropic
documents three crawlers:
ClaudeBot for training, Claude-SearchBot for
search result quality, and Claude-User for visiting sites when a person asks a question. Its
bots honor standard robots.txt directives. The interesting part comes earlier. In Anthropic's
web search documentation,
Claude "determines when to search based on the prompt," searching when the request depends
on information that is current or changing, and answering directly when the request draws
on stable knowledge.
For a local business, that split usually falls your way. Which roofer to call in a specific
town in 2026 is current information, so the search happens. And when it does, citations are
always enabled, each one carrying the source URL, the page title and a short quoted passage.
Your name arrives attached to a link rather than absorbed anonymously into a paragraph.
The part that should give you pause is the other half. When an assistant answers from what
it already believes rather than from a fresh search, the accuracy of that belief is the
whole game. Entity hardening, which means removing the gray areas so a machine can tell
exactly what your business is and what it does not do, is the work that fixes it.
Ask us about that clarity; most owners have never had anyone build it.
What works the same across all four?
Here is the finding the agency internet keeps burying. The four assistants converge much
more than they diverge. All of them want a page they can reach, an answer near the top of
it, and a business they can identify with confidence. Nobody publishes a preference for a
listicle over an essay, or for 1,200 words over 2,000. Those claims come from vendors
selling a tracking subscription, not from the companies building the systems.
Four things pay on every one of them. Let the right bots in, which means reading your own
robots.txt instead of assuming. Answer the question in the first paragraph of the page,
because a machine lifting an answer takes the first complete one it finds. Write facts as
standalone sentences that survive being quoted without the paragraph around them, since
that is literally what a citation is. And describe your business identically everywhere it
appears online, so the same name, address, services and service area show up in every
place a model might check. That last one is the least exciting and does the most work.
If you are holding a quote from another agency that prices Perplexity, Gemini and Claude as separate projects, the worry is fair: you may be paying several times for one job. In practice, most of it is one job. People call this Perplexity SEO, Gemini SEO or AI assistant SEO depending on who is
selling. The vocabulary is unsettled and the underlying job is not. Anyone quoting you four
separate content strategies is quoting three you do not need. One well-built site, checked
per assistant for access, covers the field.
That combination of access work and entity clarity is what the GetLocalLeads.AI
AI visibility and brand management
service exists to handle.
How do you check whether each assistant can see you?
Open yourdomain.com/robots.txt in a browser. Read it yourself, even if you do not consider
yourself technical, because you are looking for names rather than syntax. The ones that
matter are GPTBot, OAI-SearchBot and ChatGPT-User for OpenAI, PerplexityBot and
Perplexity-User for Perplexity, ClaudeBot, Claude-User and Claude-SearchBot for Anthropic,
and Google-Extended for Gemini training. A Disallow line under any of those names is a
decision somebody made on your behalf. Our explainer on
robots.txt and llms.txt covers what
belongs in the file.
After robots.txt, I check that llms.txt exists and actually serves the file; llms.txt is a short plain text file that AI tools read to learn about a business, and I treat it as important. Then I confirm the sitemap, the file that lists your pages for crawlers, is declared and lists real URLs, and that the business writes its name the same way on its site and its profiles. I would fix all four before anything else, because they are the cheapest to check and the most often broken.
Then run the test that costs nothing. Ask each of the four assistants the question your
customer would ask, in your own city, and write down the names that come back. Ask more than once before you believe the result: we treat a recommendation as real only when it holds across several engines and repeated runs, because one good answer proves very little. Four rows,
one per assistant, plus a column for whether you appeared at all and a column for who
appeared instead. The gaps in that little table tell you where to look first, and they are
usually not where owners expect. An owner who is named in ChatGPT and missing from
Perplexity has an access problem, not a content problem.
Changes here can move quickly. AI visibility moves faster than traditional SEO, and
movement is possible within 24 hours of a website change such as an updated robots.txt or
llms.txt file, though results are not guaranteed. Our
eight-week case study
shows what a fuller push looks like.
The honest limit is that AI answers change from one run to the next and from one engine to the next. Even the market leaders in those gap reports missed a quarter or more of the answers, which is why we call faster movement possible, not guaranteed. Fixing robots.txt gives every assistant the chance to read you; it cannot decide which names any of them says next week.
The done-for-you version of that check is our free AI visibility audit.
Frequently asked questions
Does Perplexity use Google's index?
Perplexity runs its own crawler, PerplexityBot, whose documented job is surfacing and
linking websites in Perplexity's search results, and which is explicitly not used to train
foundation models. The documentation stops there, so anyone naming the index underneath is
guessing. People ask how to rank on Perplexity, which is the wrong verb: it cites sources
rather than ranking them.
Should I block AI crawlers from my website?
Rarely, and never in bulk. All four separate training from search, so you can refuse one and
allow the other. OpenAI's documentation notes that a webmaster can allow
OAI-SearchBot while disallowing GPTBot, and that sites opted out of OAI-SearchBot "will not
be shown in ChatGPT search answers." A blanket block removes a business from answers while
failing at its purpose.
Is Gemini the same as Google AI Overviews?
Related but not identical. Gemini models power Google's AI features, and Google documents
one eligibility rule for those surfaces: a page must be indexed and eligible to show with a
snippet in Google Search. Google-Extended is a training control and does not affect Search
inclusion.
Does Claude actually read my website?
It can. Anthropic runs Claude-User, which visits sites when a person asks, and
Claude-SearchBot, which works on search quality. Claude decides per question whether to
search, answering from existing knowledge on stable topics and searching when the
information is current, as it is for which local business to hire.
Which AI assistant should a local business focus on first?
Start where the volume is: by Pew's February 2026 numbers that is ChatGPT at 44% of U.S.
adults and Gemini at 24%. Sequence matters less than people think, because the work that
earns a mention in one earns it in the others. Get your access right and your business
described consistently, then verify each one. If you would rather someone ran it, book a
call.
Where to start this week
Do the two cheap things first. Read your robots.txt and note every AI user agent named in
it, then ask all four assistants your customer's question and record who they name. Most
owners find a surprise in the first task and an uncomfortable answer in the second.
Perplexity optimization, and the same work for Gemini and Claude, rarely starts with writing
more content. It starts with finding out what you have already been blocking. When you want
the whole picture rather than two data points, we will handle the brand management
side for you.