What is answer engine optimization?
The short answer: it is the work of earning a place in the answers AI systems write. An answer names a few brands and leaves the rest out; answer engine optimization is everything that makes yours more likely to be among them, described correctly, with your own website cited.
An answer engine is an AI system that replies with a written answer instead of a list of links. ChatGPT, Claude, Gemini, Perplexity, Grok and DeepSeek all work this way, and search engines such as Google now show written answers of their own in their results. [2] A buyer who asks which 65-inch TV suits a bright living room gets one composed reply with a handful of brands in it.
That changes the work, because a written answer has no second page. In a list of links, a buyer can scroll past ten results and still find you further down; in an answer, either you are named or you are not, and the buyer may never open the sources behind it. People also click less when a summary answers them first: in a Pew Research Center study, Google users who saw an AI summary above their results were less likely to click a link than those who did not. [3] So the work is judged by how often answers name you, not by where a page ranks.
Two names for one discipline
You will also meet the name generative engine optimization, or GEO. It comes from a 2023 research paper [4], and it describes the same work. This guide says AEO throughout, and mentions GEO only here.
AEO and SEO
AEO does not replace search engine optimization: it is built on it. Pages still have to be crawled, indexed and worth reading, and the same search systems still decide which pages exist for an answer to draw on. What changes is what comes back, and so what success looks like and how you measure it.
| Compared | Classic search | Answer engine |
|---|---|---|
| What comes back | A ranked list of pages | A written answer that names brands, sometimes with sources |
| What success means | Your position | Being named, how prominently, whether you are described accurately, whether your site is cited |
| Stability | Positions drift | Answers vary from one run to the next, even for the same question |
| How to measure | Rank tracking | Shares of many answers, with a range |
| How fast things change | Crawl and re-rank | Days on the search route; model releases on the memory route |
What a brand wants from an answer
A brand wants three things from an answer: to be named, to be described accurately, and to have its own website cited. aSERP measures two of them directly. Being named is Visibility, with Prominence and Share of answer for how prominently you are placed; having your website cited is the citation rate. Whether the description is accurate, you check by reading what each engine answered: the Topics view shows the answers themselves. aSERP does not score accuracy, so reading the answers stays part of the work. An accurate description matters as much as the mention: an answer can name you and still get your prices, your range or your delivery area wrong.
Next: Where do AI answers come from?
Where do AI answers come from?
The short answer: an answer engine names brands either from what its model learned in training, or from pages it finds by searching the web while it answers. Many engines can do both, and what helps you depends on the route.
1From the model's memory
- reference sites, press coverage, reviews and video
- a language model, trained on a very large body of text
- an answer from learned associations alone
this route moves slowly, usually when the provider releases a new model
2Search, then answer
- the question
- one or more searches (query fan-out)
- results from an index
- a few picked
- the answer, written from them
retrieved is not cited
Route 1: from the model's memory
A language model is trained on a very large body of text, and along the way it learns which brands go with which needs. Ask a model that isn't searching "Where should I buy a 65-inch TV online?" and it answers from those learned associations alone. It reflects what was written about you before its training data was gathered, and mostly what was written widely and often. So this route moves slowly, usually when the provider releases a new model.
What helps here is being described often, consistently and accurately in the places models learn from: reference sites, press coverage, reviews and video.
Route 2: search, then answer
Engines that can search decide, question by question, whether the web will improve the answer. Google's documentation for grounding Gemini in Google Search puts it plainly: "The model analyzes the prompt and determines if a Google Search can improve the answer." [5] Anthropic's help center says Claude "searches the web when it helps." [6] When an engine does search, it turns the question into one or more searches (when it runs several related searches at once, Google calls this query fan-out [2]), pulls results from an index, picks a few, and writes the answer from them.
Two details matter more than most advice admits.
- The engine often reads very little of your page. In an analysis of 1,249 ChatGPT answers that searched the web (July 2026), the model usually worked from a title, an address and a snippet of roughly two hundred characters. It opened barely more than one retrieved page in 80, almost always in its thinking mode. [7: a single agency study based on reverse engineering] The first sentence under a clear heading may be all it sees of you.
- Retrieved is not cited. The same analysis followed about 58,000 retrieved addresses down to about 5,000 cited in the text of answers. [7] In controlled tests across six models and 252,000 trials, the biggest drivers of which source got cited first were its relevance to the question and how early it appeared in the material the model was given. An explicit price and a recent date helped; formatting-only edits made little difference. [8: a preprint; a laboratory setting]
Which route was a given answer on?
Apps often show that they searched (a list of sources, or a "searching the web" indicator), but not which parts of the answer came from those pages and which from what the model already knew, and they change their behaviour often. That is one reason answers to the same question can name different brands on different days, and why a single screenshot proves very little. See How do you measure visibility in AI answers?
What this means for you
- For the memory route: be described widely, consistently and accurately: earned coverage, real reviews, reference pages, and one consistent brand name. Expect it to move at the pace of model releases.
- For the search route: let each engine's search crawler reach you, rank in the indexes these engines draw on, and put the answer, with concrete facts and a real date, in the first lines of the page that matches the question.
- On both routes: be mentioned widely. In one study of 75,000 brands, how often a brand was mentioned on the web and on YouTube tracked its visibility in ChatGPT and in Google's AI features (features that search before they answer) far more closely than the number of pages on its own site did. [9: a vendor study; a correlation, not proof of cause]
Next: How do answer engines decide which brands to name?
How do answer engines decide which brands to name?
The short answer: there is no hidden switch. Engines name brands that are relevant to the exact question, present in sources the engines already trust, described with concrete and current facts, and consistently identified. The list changes from one run to the next, around a core of brands that keep coming back.
Six forces shape that list. The studies behind them differ in kind and strength, so read each card for what it is: the strength of the evidence behind one force, the route it works on, and where you can watch it in aSERP.
Relevance to the exact question
The page that answers the buyer's actual question is the one most likely to be picked. In 252,000 controlled trials, relevance to the question was the largest driver of which source was cited first [8], and a survey of 45 studies found it the most reproducible lever in the field. [10: a preprint]
What to do. Answer the question your buyers actually ask, with the comparisons and the follow-up questions that come with it. A page that answers in its first lines gives an engine something to quote even when it reads only a snippet.
In aSERP. Citations›Winnable answers lists the answers you could win, each with the next step for a page; the Content Opportunity recommendation flags the gaps; and Visibility by topic shows where you are missing.
Presence in sources engines already trust
Press, reviews, reference sites and video feed both routes, and the memory route most of all. A public-relations firm found that 84% of more than 25 million links cited in AI answers were earned media [11: a vendor study]; a research paper found a "systematic and overwhelming bias towards earned media" [12: a preprint]; and across 75,000 brands, how widely a brand was mentioned tracked how often answers named it. [9]
What to do. Earn coverage where your buyers' questions are discussed. It is slow, and it feeds both routes.
In aSERP. The News & Media Opportunity, Partnership Opportunity and Source Opportunity recommendations, and the sources answers cite most, under Citations.
Concrete, attributable facts
Prices, specifications, dates and definitions give an engine something to quote. In laboratory tests, an explicit price and a recent date helped a source get cited [8], and studies of the pages answers cite find them rich in definitions and numbers. [13: a preprint] [4]
What to do. State facts plainly, where a reader would look for them, and show when they were checked.
In aSERP. The Content Improvement recommendation.
Real freshness
Answer engines lean towards recent pages. Pages cited by AI assistants were about 25.7% fresher than the pages in ordinary search results [14: a vendor study], and in a laboratory study, systems that re-rank search results favoured text that only looked newer. [15: a preprint] Changing a date without changing the page is manipulation, and engines are building defences against it. [16] [17] [18]
What to do. Update a page when its facts change, and show the real date. Never re-date a page that did not change.
In aSERP. The note above your own pages under Citations: “These are your winning pages — keep them fresh, and replicate their structure for topics where you have zero citations.”
Entity clarity: one name, consistent facts
A brand known by several names, or described with conflicting facts, is harder for an engine to recognise and easier to confuse with another. The evidence here is how the systems match names, not a study of this force alone, and it is why aSERP's own free audit asks for every name a brand goes by: “Engines name you with whatever spelling they learned.”
What to do. Use one official name, list its other spellings everywhere you control, and keep your core facts the same on every page and profile.
In aSERP. The Entity Clarity recommendation, and the Also known as field on each competitor, where every alias you add is counted as that brand.
Access: let the crawlers in
An engine's search can only use a page that the crawlers it relies on can fetch. OpenAI, Perplexity and Anthropic document their crawlers [19] [20] [21], and the engine table in the next chapter shows what each engine documents; Cloudflare began blocking AI crawlers by default on new domains in 2025 [22] [23]; and in 2024, most of the AI crawlers studied did not run JavaScript, so facts that appear only through scripts may never be read. [24: dated evidence]
What to do. Allow the search crawlers you want to be cited by, check your firewall as well as robots.txt, and put your key facts in the page's HTML. Blocking a search crawler removes you from that engine's search; allowing it does nothing on its own.
In aSERP. The Crawlability Check recommendation.
A core of brands that keep coming back
Across repeated runs of the same prompt, ChatGPT and Google's AI Overviews Not tracked by aSERP returned the same list of brands less than 1% of the time. Yet across 994 answers to differently worded questions about one need, the same four brands each appeared in 55 to 77% of the answers. [25: a vendor-partner study reported by Search Engine Journal, with data from November and December 2025] The lesson: aim to be in the core that keeps coming back for your questions, and measure how often you are named, not what rank you hold.
Read the six forces together. Relevance, concrete facts, freshness and access matter most when an engine searches; presence in trusted sources and one clear name are associated with being named, on either route. A brand that is easy to find but rarely mentioned elsewhere can still be left out, and a brand that is widely discussed but blocks the crawlers can only be named from what a model learned.
Next: Which answer engines matter, and how do they differ?
Which answer engines matter, and how do they differ?
The short answer: aSERP tracks six answer engines: ChatGPT, Claude, Gemini, Perplexity, Grok and DeepSeek. Google's AI Overviews Not tracked by aSERP and AI Mode Not tracked by aSERP, Microsoft Copilot Not tracked by aSERP and Meta AI Not tracked by aSERP matter to buyers too, and this guide marks them wherever they appear.
The six differ in where their answers come from. Some answer mostly from what their model learned; some search their own index or a partner's; most can do both. For you, the practical difference is the crawler: where an engine documents a search crawler, that is the one you allow so it can find your pages. Google's AI features in its search results draw on the same index and the same crawler as its search [2], so what you do for Google Search reaches them too, although aSERP does not track them.
The table below says only what is public about each engine, from each company's own documentation where it exists. A search crawler and a training crawler are separate decisions: allowing one does not allow the other. Each row carries the day its facts were last checked at their sources.
| Engine | Where its answers come from | Search crawler: allow it to be found | Training crawler: a separate decision |
|---|---|---|---|
 ChatGPTChecked on 10 Oct 2026 | Model memory; when it searches, OpenAI's own index and its partners' indexes[7][26] | OAI-SearchBot[19] | GPTBot |
 ClaudeChecked on 10 Oct 2026 | Model memory; searches the web when it helps[6] | Claude-SearchBot[21] | ClaudeBot |
 GeminiChecked on 10 Oct 2026 | Model memory; can ground its answers in Google Search[5] | Googlebot | Google-Extended, a token in robots.txt rather than a crawler: it is how you decide whether Gemini may use your pages, for training and for grounding its answers; it does not affect Google Search[27] |
 PerplexityChecked on 10 Oct 2026 | Its own search index, filled by its own crawler[20] | PerplexityBot | Not used for training[20] |
 GrokChecked on 10 Oct 2026 | Model memory; tools to search the web and X[28] | No documented crawler | None documented |
 DeepSeekChecked on 10 Oct 2026 | Model memory; its app can search the web through outside services[29] | No documented crawler | None documented |
| AI Overviews Not tracked by aSERP and AI Mode Not tracked by aSERPChecked on 10 Oct 2026 | The Google Search index, with query fan-out[2] | Googlebot | Google-Extended does not affect Search[27] |
| Microsoft Copilot Not tracked by aSERPChecked on 10 Oct 2026 | Bing's index[30][31] | Bingbot, with IndexNow | None documented |
| Meta AI Not tracked by aSERPChecked on 10 Oct 2026 | Its own index, which Meta is building[32] | Meta-WebIndexer | Meta-ExternalAgent |
Using the table
Start with the engines your buyers actually use, then work through each row.
- Allow the search crawlers. In your robots.txt, allow the search crawler of each engine you want to be found by, such as
OAI-SearchBot,Claude-SearchBot,PerplexityBotandGooglebot. Blocking a search crawler takes your pages out of that engine's search, whatever else you do. - Decide on the training crawlers separately.
GPTBotandClaudeBotcollect pages for training models, not for an engine's search [19] [21], so whether you allow them does not change whether the search crawlers can reach you.Google-Extendedis different: it is a token in robots.txt, not a crawler of its own, and it does not affect Google Search, but it is how you decide whether Gemini may use your pages, for training and for grounding its answers. [27] Blocking it also keeps your pages out of Gemini's grounded answers. - Check what sits in front of your site. A firewall or your network's default rules can turn a crawler away even when robots.txt allows it, so test from outside instead of only reading the file.
- Read the dates. Engines change their crawlers and their documentation, so each row says when its facts were last checked at the source.
Two traps. A fetcher that acts for a person who asked a question can ignore robots.txt: OpenAI says of ChatGPT-User that "robots.txt rules may not apply" [19], and Perplexity says Perplexity-User "generally ignores robots.txt rules" [20]. And a firewall can block a crawler that robots.txt allows, so check both before you conclude that you are reachable.
When the answer is a product
Answer engines are learning to sell. ChatGPT's shopping results "are not ads" [33], and its in-chat checkout gave way to checkout on merchants' own sites on 24 March 2026. [34] Microsoft's checkout inside Microsoft Copilot Not tracked by aSERP, Google's checkout in AI Mode Not tracked by aSERP and in Gemini, and Perplexity's Instant Buy all started in the US. [35] [36] [37]
Next: What works, and what is myth?
What works, and what is myth?
The short answer: the levers with the best evidence are unglamorous. Most popular AEO tricks are unnecessary for Google Search and unproven elsewhere: Google itself says you can ignore "chunking" content, "creating unnecessary AI text files (like llms.txt)" and "pursuing inauthentic mentions" for Google Search. [1]
Here is what the evidence says about the tactics most often recommended. Each carries its evidence grade: whether controlled tests, large observational studies, mixed findings or very little stand behind it.
Crawler access: robots.txt, firewall and network rules
Access prerequisite
What the guide says: Allow the search crawlers you want to be cited by, decide on training crawlers separately, and check the firewall, not only robots.txt.
Facts in the page's HTML, not only in scripts
Supported by large observational studies
What the guide says: Put the answer, the price and the availability in the HTML, in the first lines under the headline.
Relevance to the exact question
Supported by controlled tests
What the guide says: Write the page that answers the buyer's actual question.
Concrete facts: prices, dates, numbers, definitions
Supported by controlled tests
What the guide says: State them plainly, with the date they were checked.
Original data
Thin evidence
What the guide says: Supported in the laboratory, thin in the wild. Publish numbers others cannot copy, with the method behind them.
Earned media
Supported by large observational studies
What the guide says: Coverage in the publications engines already cite is the most consistent lever.
Reviews and user content
Thin evidence
What the guide says: Real reviews where buyers leave them; never fake ones. Google's spam policies now cover AI answers too. [17]
Reddit and forums
Mixed evidence
What the guide says: Take part genuinely where buyers already talk. Don't build a strategy on one platform's share: Reddit's share of ChatGPT responses swung from about 60% to about 10% within weeks in 2025. [38: a vendor study]
Wikipedia and Wikidata
Supported by large observational studies
What the guide says: Wikipedia is among the sites AI answers cite most often. [38] Make sure the facts about you are right and sourced, through Wikipedia's conflict-of-interest rules, and never write promotional articles. The evidence for Wikidata on its own is thin.
Structured data
Mixed evidence
What the guide says: Pages that are cited more often tend to carry it, but when a search-tool vendor added it to 1,885 pages, they showed no uplift against 4,000 matched pages. [39] Keep structured data accurate and matching the visible text, and do not treat it as a ranking lever.
FAQ blocks
Mixed evidence
What the guide says: Answer real questions where they belong, with no boilerplate written for machines. Google's FAQ rich result has been gone since 7 May 2026. [17]
llms.txt
Not used by Google Search
What the guide says: Optional and harmless, with no measured effect: across about 300,000 domains, having the file showed no link to being cited. [40]
"Chunking", or writing for AI
Not used by Google Search
What the guide says: Write for people. Clear headings and direct first sentences serve both.
Freshness
Supported by large observational studies
What the guide says: Update when facts change, and show the real date.
Page speed
Thin evidence
What the guide says: It matters for people and for the fetches that do happen; no study ties it to citations in AI answers.
Video
Supported by large observational studies
What the guide says: Video that names and explains your product is associated with being named in AI answers. [9]
Inauthentic mentions and spam
Harmful
What the guide says: Don't. Engines are building defences against manipulation. [16] [17]
How to read conflicting studies
A public-relations firm finds that 84% of the links AI answers cite are earned media [11]; a business-listings firm finds that 86% are brand-managed. [41: a vendor study] Both can be true. Ask four questions of any statistic before you act on it: which engines, which questions, which period, and who sells what.
If you can do only three things, do the ones with the strongest evidence: make sure the search crawlers can reach your pages, answer your buyers' exact questions on pages of their own, and put concrete, dated facts in the first lines. Everything else in the list has weaker evidence behind it, or is not used by Google Search at all.
Next: How do you measure visibility in AI answers?
How do you measure visibility in AI answers?
The short answer: as a share of many answers over measured days, with a range. Never as a single answer, and never as a rank.
Why one answer is not a measurement
A single answer is one day, on one engine. Ask the same question again and the list of brands changes: across repeated runs of the same prompt, ChatGPT and Google's AI Overviews Not tracked by aSERP returned the same list less than 1% of the time. [25] Even with the setting meant to make answers repeatable, 1,000 identical requests to one open model on a standard setup gave 80 different completions. [42] This is non-determinism, and it is why a figure has to be a pattern across many answers.
A share is added up across answers: the answers that name you, divided by all the answers measured in the period. It is never an average of daily percentages, so a day with more measured answers counts for more than a day with fewer.
Five figures, one line each
aSERP reports five figures for each brand. Each has a full entry in the Glossary.
- Visibility: the share of measured answers that name you, with a 95% range. Answers to the same question move together, so read it with the number of questions behind it and its confidence label.
- Prominence: how early and how strongly you are placed in the answers that name you, on a 0–100 scale.
- Share of answer: your share of the brands aSERP identifies in each answer, weighted by how prominently each is placed, averaged over the measured answers.
- Citation rate: of the answers that name any website, the share that name or link yours. For every engine, aSERP counts a website only when the answer's own text names or links it.
- Answer stability: how steady the mix of brands is from one measured day to the next.
Ranges and confidence labels
Every figure carries a confidence label: High confidence, Directional or Low confidence. Visibility and Share of answer also show their 95% range, on the Competitors chart. The reader's rule: trust a direction only when the ranges stop overlapping, on comparable periods measured the same way. Two ranges that still overlap are compatible with no change at all, however different the two figures look.
One caution about ranges. A range computed over answers does not widen when the answers come from only a few questions, and answers to the same question tend to move together, so such a range can look more certain than it is. Compare periods measured on the same set of questions, with enough questions in each topic.
Missing is not zero
When an engine kept failing to answer on a day, aSERP marks it AI unavailable, and those answers are not counted as "not mentioned". A gap in the measurement is not a sign that you were left out. A value that cannot be computed shows as a dash, never as 0.
A figure belongs to the days it was measured
Every figure belongs to its measured days. aSERP compares a period with the equal period just before it, measured the same way, and declines a comparison when the method changed in between or too few days were measured.
Why your screenshot and our figure can disagree
Our figures are our own measurement, not what any one person sees in an app, where memory of earlier chats, location, personalization and model versions can differ. [43] [44] [45] [46] Answers also vary from run to run, as above. aSERP asks six answer engines the same questions, the same way, every day; a screenshot is one answer, on one day, for one person.
Every vendor's metric means something else
Read every tool's definition before you compare numbers across tools. Bing's "Citation Share" is "not a ranking system" [47], and Google's report on AI features in Search Console counts impressions. [48] Neither is the same thing as how often answers name your brand. A share of voice, a citation share and a visibility score can each be honest and still measure different things.
The free audit is a first look
The free audit asks three of your buyers' questions on six engines, once: a snapshot, not a trend. Its score is deliberately conservative and capped well below 100, and it is not the Visibility figure.
A first look, not a trend. The free audit asks six engines three of your buyers' questions, once. Use it to see where you stand today, and daily tracking to see where you are heading.
Run a free auditNext: How do you turn measurement into action?
How do you turn measurement into action?
The short answer: in a loop. Choose the questions buyers ask; measure; find the answer gaps; change one thing; measure the same questions again over comparable periods; and judge with ranges, as an observation, not as proof of cause.
Start from the questions
Start from the questions your buyers ask, by intent: the free audit asks one question to learn, one to compare and one to buy. Group them into topics, and add as competitors the brands that are named instead of you. For a TV retailer, a question to learn might be "What should I look for in a TV for a bright room?", a question to compare "Which is better for a bright room, an OLED or a mini-LED TV?", and a question to buy "Where can I buy a 65-inch TV with free delivery this week?" Write the questions your buyers actually ask AI shows how to write the questions, and Choose the competitors to track how to pick the competitors.
Then change one thing at a time. If you rewrite a page, earn new coverage and rename a product in the same week, no comparison can tell you which of them moved the answer, and some of them may have cancelled each other out.
Find the answer gaps
An answer gap is a question where competitors are named and you are not. aSERP turns the gaps it finds into recommendations, in three areas, as the product groups them.
On-page
- Citation Opportunity
- Content Opportunity
- Content Improvement
- Crawlability Check
Off-page
- Partnership Opportunity
- Source Opportunity
- Reddit Conversation
- News & Media Opportunity
- Video Source Opportunity
- Social Conversation
General
- Competitive Visibility
- Entity Clarity
- Prompt Strategy
- Branded Prompts
- Awareness-stage Prompt Coverage
- Brand Visibility
How aSERP judges a recommendation
After you act on a recommendation, aSERP compares a baseline period with a later one. It reports Observed improvement only when the change is larger than the metric's minimum change or the two periods' uncertainty added together, whichever is greater; otherwise No material change or Declined, and Insufficient support when there is too little data. The Experiments view shows the averages before, during and after a change: read it with the ranges and the coverage, not as a verdict on its own. Act on a recommendation and check what changed walks through one recommendation from start to finish.
What you cannot conclude
A comparison of two periods can show that a figure moved after a change, not that the change moved it: correlation is not causation, and only a controlled test can show the second. In the product's own words: “Before/after comparisons are observational, not causal attribution.”
Engines move at different speeds
The memory route reflects a change only after a model update; the search route can within days, once the engine has read the page again. Read each engine separately with the Platform filter, and expect that not every engine will move. No movement on one engine does not mean the change failed on the others.
aSERP measures the same questions every day across six engines, turns what it finds into recommendations, and lets you log each change as an experiment with before-and-after periods. See plans.
Start freeNext: Where to start
Where to start
The short answer: nine steps, in this order. Where a step has its own How-to guide, the guide is named after it.
- Write down the questions your buyers ask an AI, in their words, by intent. Guide: Write the questions your buyers actually ask AI.
- Check that the search crawlers you want can reach you, through robots.txt and the firewall.
- Run the free audit for a first look, and treat it as one capture. Guide: Run your first AI visibility check.
- Make your brand's names unambiguous everywhere you control.
- For each priority question, make sure one page answers it in its first lines, with concrete, dated facts.
- Earn mentions in the sources engines already cite for those questions.
- Keep facts current, and change a page's date only when its content changes.
- Measure across engines over many days, and read shares with ranges, never single answers. Guide: Read your AI visibility numbers without fooling yourself.
- Change one thing at a time, compare comparable periods, and write down what changed and when.
The order matters. The questions come first because every later step is judged against them; access comes second because a page the engines cannot fetch cannot be found by their search; and measurement comes after the changes, over enough days, because a change shows on different engines at different speeds.
Next: Questions people ask
Questions people ask
The short answer: these are the questions marketers ask most about answer engines, each answered in a few lines, with the chapter that goes deeper.
Is AEO different from SEO?
It is built on SEO. Google's own guidance says that optimizing for its AI features "is optimizing for the search experience, and thus still SEO." [1] What differs is the result you aim for: a place in a written answer that names a few brands, which you measure as a share of many answers rather than as a rank. See What is answer engine optimization?
Can I pay to be named in AI answers?
You can buy ads that appear beside some answers. ChatGPT has shown labelled ads on its free and lowest-priced tiers in some countries since February 2026. [49] OpenAI's stated rule is that ads do not influence the answer itself [50], and its shopping results are chosen independently and are not ads. [33] Google also shows labelled "Sponsored" ads within some AI Overviews Not tracked by aSERP, in the US and several other countries. [51] [52] In both cases the ad is marked and separate from the written answer: there is no paid way into the answer's own text.
How long until a change shows up?
It depends on the route. On the memory route, a change shows only after the provider trains and releases a new model. On the search route, the engine must crawl and index the page again first. The one published timing is for access rules: OpenAI says its systems can take about 24 hours to adjust after a robots.txt change. [19] Content changes have no published timetable, so measure daily and compare equal periods. Engines differ in how fast they can reflect a change: read each engine separately, and expect that not every engine will move.
Does llms.txt help?
No measured effect [40], and Google Search does not use it. [1] An llms.txt file is optional and harmless; spend the time on the page that answers your buyers' question instead.
Does structured data help?
The evidence is mixed. Pages that are cited more often tend to carry it, but when a vendor added it to pages and compared them with matched pages, it showed no uplift. [39] Keep structured data accurate and matching the visible text, and do not expect it to place you in answers on its own.
Why do your figures differ from what I see in ChatGPT?
Our figures are our own daily measurement, not what any one person sees in an app, where memory of earlier chats, location, personalization and model versions can differ; and answers vary from run to run. See Why your screenshot and our figure can disagree.
Why is my free audit result different from my dashboard?
The audit is one capture: three questions, six engines, once. The dashboard measures your tracked questions over many days and reports shares with ranges. If you were new to aSERP, the audit's answers became your first brand's first measurements; if you already had an account and added the audited site as another brand, aSERP measured it again there, so the numbers can differ.
Which engines does aSERP track?
Six: ChatGPT, Claude, Gemini, Perplexity, Grok and DeepSeek. The engines aSERP tracks do not include AI Overviews Not tracked by aSERP, AI Mode Not tracked by aSERP, Microsoft Copilot Not tracked by aSERP or Meta AI Not tracked by aSERP.
How many questions should I track per topic?
Enough that no single question dominates. Answers to one question move together, so ten questions asked for a month say more than one question asked three hundred times.
Next: Sources, and how this guide was made
Sources, and how this guide was made
The short answer: every outside fact in this guide links to its source, listed here with its type and the day we last checked it. Papers are listed by their titles.
- Google Search Central, “Optimizing your website for generative AI features on Google Search”.
- Google Search Central, “AI features and your website”.
- Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, 22 July 2025.
- KDD 2024 (arXiv:2311.09735), “GEO: Generative Engine Optimization”, 16 November 2023.
- Google AI for Developers, “Grounding with Google Search”.
- Claude Help Center, “Enable and use web search”.
- Resoneo, “What ChatGPT pulls, what it shows, what it cites”.
- arXiv:2605.25517, “What Gets Cited: Competitive GEO in AI Answer Engines”, 25 May 2026.
- Ahrefs, “Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews (75k Brands Studied)”.
- arXiv:2607.14035, “Optimizing Visibility in Generative Engines: A Critical Survey of Generative Engine Optimization (2023-2026)”, 15 July 2026.
- Muck Rack Generative Pulse, “What Is AI Reading?”, 7 May 2026.
- arXiv:2509.08919, “Generative Engine Optimization: How to Dominate AI Search”, 10 September 2025.
- arXiv:2604.25707, “From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms”, 28 April 2026.
- Ahrefs, “Do AI assistants prefer to cite fresh content?”, 28 July 2025.
- arXiv:2509.11353, “Do Large Language Models Favor Recent Content? A Study on Recency Bias in LLM-Based Reranking”, 14 September 2025.
- arXiv:2609.02964, “When Optimization Becomes Manipulation: Defending Generative Search against Malicious Generative Engine Optimization”, 2 September 2026.
- Google Search Central, “Latest documentation updates”.
- Google Search Central, “Creating helpful, reliable, people-first content”.
- OpenAI, “Overview of OpenAI Crawlers”.
- Perplexity, “Perplexity Crawlers”.
- Claude Help Center, “Does Anthropic crawl data from the web, and how can site owners block the crawler?”.
- MIT Technology Review, “Cloudflare will now, by default, block AI bots from crawling its clients' websites”, 1 July 2025.
- Cloudflare, “Cloudflare Just Changed How AI Crawlers Scrape the Internet-at-Large; Permission-Based Approach Makes Way for A New Business Model”, 1 July 2025.
- Vercel and MERJ, “The rise of the AI crawler”, 17 December 2024.
- Search Engine Journal, reporting a SparkToro and Gumshoe study, “AI Recommendations Change With Nearly Every Query: SparkToro”, 30 January 2026.
- Falia, “What search engine does ChatGPT use?”, 8 September 2026.
- Google Search Central, “List of Google's common crawlers”.
- xAI documentation, “Web Search”.
- DeepSeek, “DeepSeek Privacy Policy”.
- Bing Search Blog, “Evolving role of the index: From ranking pages to supporting answers”, 6 May 2026.
- Bing Webmaster Tools, “Webmaster Guidelines”.
- PPC Land, “Explaining Meta-WebIndexer”, 29 August 2026.
- Search Engine Land, “OpenAI adds shopping features to ChatGPT Search”, 28 April 2025.
- CNBC, “OpenAI revamps shopping experience in ChatGPT after Instant Checkout”, 24 March 2026.
- GeekWire, “Microsoft debuts Copilot Checkout”, 8 January 2026.
- commercetools, “Get ready for agentic commerce: A merchant's guide to Google's Universal Commerce Protocol (UCP)”, 3 June 2026.
- PYMNTS, “PayPal and Perplexity Debut Instant Buy and 50% Off”, 25 November 2025.
- Semrush, “The Most-Cited Domains in AI: A 3-Month Study”, 10 November 2025.
- Ahrefs, “Schema and AI citations study (1,885 treated pages, 4,000 controls)”, 11 May 2026.
- SE Ranking, “llms.txt study (about 300,000 domains)”, 7 November 2025.
- Yext, “AI Doesn't Rank, It Cites. And 86% of Its Sources Are Brand-Managed”, 9 October 2025.
- Thinking Machines Lab, “Defeating Nondeterminism in LLM Inference”, 10 September 2025.
- TechCrunch, “ChatGPT will now use its 'memory' to personalize web searches”, 18 April 2025.
- Search Engine Land, “ChatGPT enables location sharing for more precise local responses”, 31 March 2026.
- Google, “Personal Intelligence: Connecting Gemini to Google apps”, 14 January 2026.
- Gemini Apps Help, “Gemini Apps limits & upgrades for Google AI subscribers”.
- Bing Search Blog, “New AI Visibility Insights in Bing Webmaster Tools: Intents, Topics, Citation Share, Compare”, 16 June 2026.
- Google Search Central Blog, “Introducing Search Generative AI performance reports in Search Console”.
- TechCrunch, “OpenAI to start showing ads on ChatGPT's free and Go tiers in India”, 27 August 2026.
- OpenAI Help Center, “Ads in ChatGPT”.
- Google Ads Help, “About ads and AI Overviews”.
- Google, “New ways for marketers to reach customers with AI Overviews and Lens”, 3 October 2024.
How this guide was made
This guide is prepared and checked by aSERP Research. A writer is responsible for it; a second reviewer checks every claim against its source; whoever owns our metrics signs off every sentence about how aSERP measures. Each tactic carries an evidence grade, and each outside fact its source and the day we checked it.
Sources are chosen in a fixed order of preference: official documentation first, then peer-reviewed papers, then preprints (papers not yet reviewed), then independent research, then vendor studies, which are labelled as such, and press last. Every outside source is opened again before the guide is published or changed, and the day of that check is printed in the list. When two studies disagree, the guide says so instead of choosing the more convenient one. A grade describes the strength of the evidence behind a tactic, not how much the tactic is worth to your brand.
It covers how answer engines choose the brands they name and how to measure your place in their answers. It does not cover paid advertising beyond the question above, and it treats shopping inside answer engines only briefly. It does not rank engines or brands, and it names no business that our research measures. When we get something wrong, we fix it and add a dated note that says what changed. Spotted an error? Tell us.