Cara Mengoptimalkan Situs Web Anda untuk Pencarian Berbasis AI

If you sell anything online, you need to optimise your website for AI-driven searches now rather than later. Shoppers increasingly start with a chat assistant or an AI summary instead of a list of ten blue links, and the page that gets quoted in that answer is the page that earns the visit. The good news is that this is not a new discipline bolted on top of SEO. It is mostly SEO done properly, plus a handful of genuinely new controls.

The traffic shift is measurable. Adobe Analytics found that traffic from generative AI sources to U.S. retail sites rose 1,300% year over year between 1 November and 31 December 2024, then 4,700% year over year in July 2025. Growth has since moderated to 769% year over year in November 2025 and 673% in December 2025 — smaller multiples off a much bigger base. In Adobe's survey work, 38% of consumers said they had already used generative AI for online shopping.

What makes that traffic worth chasing is its quality. Over the 2025 holiday season Adobe reported that visitors arriving from AI sources converted 31% more often than visitors from other sources, spent 45% more time on site and were 33% less likely to bounce. Meanwhile Pew Research Center found that when Google showed an AI summary, users clicked a traditional result in 8% of visits versus 15% when no summary appeared. Fewer clicks, better clicks. This guide covers what the primary sources actually say you should do — and what they say you can skip.

Separate what is documented from what is guesswork

A large share of "AEO" and "GEO" advice circulating online is speculation dressed up as fact. Before you spend a sprint on it, check it against what Google, Microsoft and OpenAI have actually published. The table below contrasts the most common claims with the official position as documented at the time of writing.

Common claim	What the primary sources say	Verdict
You must publish an llms.txt file to be cited by AI	Google's guide to optimising for generative AI features states you do not need to create new machine-readable files, AI text files, markup or Markdown, and names llms.txt specifically. No major search engine has announced support for it.	Unsupported
There is special schema for AI Overviews	Google's AI features documentation says there is no special schema.org structured data needed to appear, and that structured data is not required for generative AI search. Structured data still earns rich results in classic Search.	Myth (but schema is still worth doing)
You need to chunk content into tiny standalone blocks	Google states it understands nuanced pages that cover several subtopics and that content does not need breaking into small pieces.	Unsupported
You should write differently for machines than for people	Google says systems understand synonyms and varied phrasing, so writing a separate AI-facing version is unnecessary.	Unsupported
Buying brand mentions across the web gets you cited	Google explicitly calls out seeking inauthentic mentions as ineffective.	Risky, avoid
Being indexed is the prerequisite for AI visibility	Google's documentation is explicit: pages must be indexed and eligible to appear with a snippet. There are no additional requirements beyond that.	Documented
You can opt out of AI answers	Google exposes preview controls (nosnippet, data-nosnippet, max-snippet) plus a Search Console generative AI toggle; opting out forfeits traffic and impressions from those features.	Documented

Treat anything outside that second column as a hypothesis. Some of it may prove right — the field is young — but present it as an experiment with a measurement plan, not as an established ranking factor.

Make sure the right crawlers can reach you

Visibility in an AI assistant depends on a bot being allowed to fetch your pages. The bots are not interchangeable, and blocking the wrong one quietly removes you from answers.

Crawler	Operator	What it does	Effect of blocking it
OAI-SearchBot	OpenAI	Surfaces sites in ChatGPT search	You drop out of ChatGPT search answers
GPTBot	OpenAI	Collects content for model training	Content excluded from training; search visibility is unaffected
ChatGPT-User	OpenAI	Fetches a page when a user or a GPT action asks for it	User-triggered fetches may fail
Googlebot	Google	Standard crawling and indexing, which AI Overviews and AI Mode build on	No indexing, therefore no AI feature eligibility
Google-Extended	Google	Controls use of your content for Gemini model training and grounding	Training use restricted; Search indexing unaffected
Bingbot	Microsoft	Indexing that underpins Bing and Copilot answers	No Bing or Copilot visibility

The most common self-inflicted wound here is a blanket rule that disallows every AI-looking user agent. If you want the referral traffic but not the training use, allow OAI-SearchBot and disallow GPTBot — OpenAI documents these as independent controls. Draft and sanity-check the file with a robots.txt generator before you deploy it, and re-test after every platform migration.

On the Microsoft side, Bing's webmaster guidance for AI-powered search is unglamorous and concrete: keep an accurate XML sitemap, reference it from robots.txt, and use IndexNow to push new and updated URLs in real time. Bing stresses that lastmod should reflect the real date the page changed, because freshness signals influence how quickly updates reach AI-generated answers. Build or repair the file with an XML sitemap generator and confirm every canonical URL is listed.

Write product descriptions the way shoppers actually ask

Assistants answer questions phrased in natural language, so the pages that answer well are the ones written in natural language. Writing product descriptions in a plain, conversational register makes them easier for a person to skim and easier for a model to reuse without distortion.

Lead with the problem the product solves, then support it with specifications. A description that opens with "keeps a 12-hour shift comfortable because the midsole is 3mm thicker at the heel" gives an assistant a claim plus its evidence; one that opens with a materials list gives it nothing to quote. Premium brands have written benefit-first for years — Apple frames iPhone copy around what the device lets you do, Gymshark around how training should feel — but you do not need their budget to copy the structure.

Three habits do most of the work:

Answer first, elaborate second. Put the direct answer in the opening sentence of a section, then add nuance.
Keep claims verifiable on the page. Sizes, materials, compatibility, warranty terms and return windows should be stated in text, not buried in an image or a PDF.
Use descriptive headings and short paragraphs. Google's own guidance asks for content organised with clear sections and descriptive headings, and for semantic HTML — while noting it does not have to be perfect.

Better copy also lifts the engagement metrics you control. It is plausible that those quality signals feed the ranking systems AI features build on, but it is not established that bounce rate is a direct AI ranking input — treat it as a mechanism, not a proven factor.

Target long-tail and question-based queries

Conversational search rewards specificity. "Waterproof hiking boots for wide feet under £120" is a query a person types into an assistant and almost never into a keyword tool. These long-tail phrases carry high purchase intent and usually face far less competition, which is why smaller sites can win them while losing the head terms.

Find them where they already exist rather than inventing them. Your customer service inbox and chat logs are the richest source: they contain the exact objections and edge cases buyers raise before paying. Complement that with the "People Also Ask" boxes on the results page and Google's autocomplete suggestions, then expand and prioritise the set with a long-tail keyword generator.

Then answer them where the decision happens: an FAQ block on the product page beats a separate FAQ hub, because it puts the answer next to the buy button. Mark it up with a FAQ schema generator if you want the rich-result upside in classic Search — just remember Google says that markup is not what gets you into AI Overviews.

Fix the technical basics: crawlability, speed and mobile

Google has used the mobile version of a site for indexing and ranking since mobile-first indexing completed, so the small-screen layout is the layout that counts. Page experience is part of Google's stated advice for generative AI features, and Core Web Vitals remain the public yardstick. The current thresholds, measured at the 75th percentile of real user page loads, are worth knowing exactly:

Metric	What it measures	"Good" threshold
Largest Contentful Paint (LCP)	Loading speed of the main content	2.5 seconds or less
Interaction to Next Paint (INP)	Responsiveness to user input	200 milliseconds or less
Cumulative Layout Shift (CLS)	Visual stability during load	0.1 or less

INP replaced First Input Delay as a stable Core Web Vital in 2024, so if your reporting still tracks FID it is out of date. Start with a crawl-level view from a SEO audit tool, then drill into templates with a page speed analyzer. Fix "Poor" scores on high-value templates first: one product template usually governs thousands of URLs.

Images are the usual culprit for LCP, and they carry a second benefit. Descriptive alt text makes image content legible to systems that parse the page as text, so run a sweep with an image alt text checker and compress oversized assets before you consider a rebuild.

If the underlying site is beyond repair, rebuilding on a modern platform is a legitimate shortcut. Hocoos notes that most AI website builders ship with built-in SEO features, and such platforms usually handle image compression, responsive layouts and clean markup by default. That saves engineering time; it does not substitute for the content and crawler work above.

A ten-step implementation plan

Confirm your key pages are indexed and snippet-eligible. If they are not in the index, no AI feature can use them.
Audit robots.txt for over-blocking. Decide deliberately, per bot, whether you want search visibility, training use, or neither.
Publish an accurate XML sitemap with honest lastmod values, reference it from robots.txt, and submit it to Google Search Console and Bing Webmaster Tools.
Turn on IndexNow so price, stock and content changes reach Bing and Copilot without waiting for a recrawl.
Mine support tickets, chat logs, "People Also Ask" and autocomplete for the twenty questions buyers ask most.
Rewrite your top twenty product or category pages answer-first, with the direct answer in the opening sentence of each section.
Add a five-to-six question FAQ to each of those pages, placed near the purchase decision rather than on a separate hub.
Add or repair structured data for products, reviews, breadcrumbs and FAQs — for rich results in classic Search, not because AI features require it.
Bring LCP, INP and CLS into the "good" band on your highest-traffic templates, starting with images.
Feed your catalogue to Google Merchant Center and keep your Google Business Profile current, which Google's own guidance names as the way to surface products and services in its AI experiences.

Measure it instead of guessing

Until recently, AI-driven visibility was largely unmeasurable, which is exactly why speculation flourished. That has changed. Google Search Console now includes generative AI performance reporting covering AI Overviews, AI Mode and AI Overviews in Discover, with impressions broken down by page, country, device and date. Note the limitation: at launch this reporting shows impressions rather than clicks, so treat it as a visibility signal, not a traffic report. Search Console also carries a control that lets a site opt out of appearing in and grounding those generative AI features, which Google rolled out globally in 2026 — and opting out means forgoing traffic and impressions from them.

Pair that with your analytics. Referrals from assistant domains arrive as ordinary referral traffic, so segment them and judge the channel on revenue rather than sessions — given Adobe's 31% conversion premium, a small session count can carry outsized revenue. Track your position on the queries that matter with a keyword rank tracker so you can tell a real shift from normal volatility.

Frequently asked questions

Do I need an llms.txt file?

No search engine has announced support for it, and Google's guidance on generative AI features says you do not need to create AI text files, naming llms.txt directly. Publishing one is cheap and harmless, but treat any claim that it improves AI visibility as speculation rather than fact.

Does schema markup get me into AI Overviews?

Not by itself. Google's documentation states there is no special structured data required for AI features and that structured data is not required for generative AI search. Schema still earns rich results in classic Search and helps disambiguate products, prices and reviews, so it remains worth implementing — just not as an AI shortcut.

If I block GPTBot, do I disappear from ChatGPT?

No. OpenAI documents GPTBot and OAI-SearchBot as separate controls: GPTBot governs model training, while OAI-SearchBot governs whether you can be surfaced in ChatGPT search. Blocking GPTBot alone does not remove you from ChatGPT search answers.

Should I write a separate version of my pages for AI?

Google advises against it, saying its systems understand synonyms and varied phrasing and that content does not need to be chunked into small pieces. Write once, clearly, with descriptive headings and semantic HTML, and judge the result on whether a human finds the answer fast.

Is AI search actually reducing my clicks?

Often, yes. Pew Research Center found users clicked a traditional result in 8% of visits where an AI summary appeared, against 15% where none did, and clicked a link inside the summary in only 1% of such visits. Plan for fewer, higher-intent sessions rather than assuming flat volume.

How do I know whether any of this is working?

Use the generative AI performance reporting in Search Console for impressions across AI Overviews and AI Mode, Bing Webmaster Tools for Copilot-side visibility, and your own analytics for assistant referrals segmented by conversion rate. Set a baseline before you change anything so you can attribute movement honestly.

Where to start this week

Optimising for AI-driven searches is less exotic than the marketing around it suggests. Be indexable, be crawlable by the bots whose traffic you want, answer real questions in plain language on the page where the purchase happens, and keep the site fast on a phone. Everything in that list also makes the site better for the humans who arrive by any route.

Pick the two changes with the shortest path to impact: audit robots.txt for accidental blocks, and rewrite the FAQ on your ten highest-revenue product pages using questions from your own support inbox. Both are a week's work and both hold up whatever the platforms change next. Add the speculative tactics later, labelled as experiments and dropped without sentiment when the data says they did nothing.