How AI Engines Answer Questions About Your Business
The mechanics behind AI search, and what they mean for your visibility
Last updated: August 2026
AI engines don’t look things up the way you do on Google. They predict what text is likely to come next, based on everything they have learned and what they can access at the moment the question is asked. That difference shapes what shows up about your business, and what doesn’t.
This page explains the mechanics. It matters because the work of being visible in AI search only makes sense once you understand how these systems behave. Otherwise, the recommendations look like a longer list of SEO tactics. They aren’t. AI Visibility is a different kind of work, aimed at a different kind of reader.
Two paths to an answer
When you ask an AI engine a question, the engine has two paths to building its answer. The question might be “what pool stores in Hartford County offer free water testing?” or “is Acme Plumbing trustworthy?”
The first path is retrieval. The engine searches for source material that addresses your question directly. That includes reviews, articles, business listings, the company’s own website, and citations on third-party platforms. If that source material is rich and specific, the engine builds its answer from it and cites the sources.
The second path is generation. The engine produces text that fits the shape of the question, drawing on patterns from millions of similar questions and answers it learned during training. It doesn’t need source material because it isn’t citing anything. It is predicting what a credible answer would look like.
Most answers are a mix. The engine retrieves what it can and generates the rest. The ratio depends on how much source material exists for the specific question.
Here is the consequence: when the source material is thin, the generated portion grows. The engine fills the gap with confident-sounding language that matches the pattern of how that kind of answer usually reads. It is not lying. It is doing what it was built to do: produce plausible continuations of text. But from your customer’s perspective, the result reads as if it came from research, even though most of it came from pattern-matching.
This is the foundational mechanic. Everything else on this page builds on it.
What AI engines pull from
The retrieval path draws from a few specific places. Knowing what they are makes it clear why our recommendations focus where they do.
- Training data is the text the model learned from when it was built. This includes a snapshot of the web up to a cutoff date, plus books, articles, and other sources. It is fixed at training time, so content published after the cutoff is not in training data.
- Live web search is how most engines now incorporate current information. The engine searches the web in real time, retrieves a small set of pages, reads them, and uses what it finds.
- Citation sources are the specific pages the engine treats as authoritative for the question. They appear as footnotes or inline links in the answer.
- Business directories and review platforms are heavily weighted because they are structured, recognizable, and updated frequently. These include Google Business Profile, Yelp, BBB, Customer Lobby, and industry-specific platforms.
- Knowledge Graph entities are the structured map of organizations, people, places, and concepts that Google maintains. Several engines reference Knowledge Graph data either directly or through the organic ranking systems they pull from.
Critically, AI engines do not distinguish well between current and historical content. A 2014 customer review and a 2025 customer review look the same to the engine if both are indexed and accessible. The engine has no reliable way to tell which one reflects current operations. It pulls from both and presents the combined picture as current. This is changing. Engines are starting to weight recency signals more heavily, but the change is partial, not complete, and it depends on whether your content has the freshness signals the engines look for.
Why this changes what “visibility” means
Traditional SEO was about ranking, being on the first page of Google for a search term. The reader still had to choose which link to click and form their own impression.
AI search compresses that step. The engine reads the sources, decides what to say about your business, and presents a summary as the answer. Most customers stop there. They do not click through to verify. The summary is the impression.
Visibility in AI search means two things at once. First, the engine has to know enough about your business to name it. Second, what the engine knows has to be accurate and current enough to represent you fairly. The first is a presence problem. The second is a source-material problem. Both have the same fix. Get more authoritative, current content about your business into the places AI engines pull from.
This is why the recommendations in our Intelligence reports look different from a traditional SEO audit. Keywords still matter, but they aren’t the lever. The lever is the volume and quality of source material available to AI engines when they get asked about your business. Producing that kind of content is what AI Visibility Content Development does.
Here is presence and source material in a single real answer:
A real result
When your content is a source, and your brand is mentioned in the answer, the engines consider you an authority on the topic.
“The Importance of STAT and Emergency MRI Services at GWIC,” and the STAT MRI prioritization process, both pulled from GWIC’s own pages.Source: greaterwaterburyimagingcenter.org
across the answer
greaterwaterburyimagingcenter.org
This is the Story of Record, working. When a business publishes authoritative content, the AI engines read it, trust it, and repeat it.
How the six major engines differ
Six AI engines currently shape what consumers find when they search for businesses, products, and services. Each has its own retrieval preferences, citation patterns, and willingness to generate when sources are thin. These differences matter for the work we recommend, because optimizing for one engine is not the same as optimizing for another. Tracking all six over time, and measuring where you gain or lose ground in each, is what AI Visibility Intelligence does.
The third column is the one to read alongside your SEO program. It describes how heavily each engine leans on your Google ranking position when it decides what to cite. The pattern that follows the table is the important part: the engines are spread across a wide range, from ones where a strong organic rank is close to a prerequisite to ones where your Google position barely predicts anything.
| Engine | Behavior | How much it depends on your Google ranking | What this means for your SEO work |
|---|---|---|---|
| ChatGPT (OpenAI) | Largest consumer AI assistant by user base. Cites sources explicitly, including Customer Lobby URLs, BBB, Reddit threads, Wikipedia, Forbes, and G2. Acknowledges uncertainty with phrases like “limited online data” rather than generating. Heavy preference for earned media in established publications. | Weak dependence on your Google rank. Retrieves through Bing’s index plus its own crawler, and studies put its overlap with Google’s top-10 results in the low double digits. Ranking #1 helps but is far from sufficient; brand mentions across authoritative sources predict citation better than position. | Citation work and earned media pay off here. Wikipedia entity links and structured data carry weight. |
| Claude (Anthropic) | Enterprise-favored for long-form reasoning. Retrieves citations primarily through Brave Search. Strong preference for primary research, methodology content, dated facts, and verifiable claims. Highest restraint on generation when sources are thin. | Lowest dependence on your Google rank of the six. It reads Brave’s index, not Google’s, so your Google position has little predictive value. Visibility on Brave and the depth and structure of your content matter more than where you rank on Google. | Original research, named methodologies, and dated case studies are what Claude rewards. |
| Gemini (Google) | Google’s flagship consumer AI assistant. Pulls heavily from a company’s own website and About page. Highest sentiment scores overall. Also higher willingness to generate plausible specifics without citation when source material is thin. | Conditional dependence. Gemini favors domains with established organic authority. Below roughly the top 20 for a query, citation odds fall off sharply, but a strong rank alone doesn’t earn a cite. It often mentions a brand from learned knowledge without linking, so passage-level clarity decides whether you’re cited or just named. | On-domain content quality directly shapes what Gemini says. Regular content refresh signals freshness to Gemini’s recency weighting. |
| Google AI Overviews | AI-generated summaries at the top of Google search results. Trigger rates vary widely by the shape of the query rather than sitting at one number. Across roughly 49,000 tracked queries, Seer Interactive measured AI Overviews on 36% of informational queries, 8% of commercial ones, and 5% of transactional ones, rising to 86% on question-format queries and 95% on comparison queries¹⁰. Citation behavior leans on high-authority organic results and established Knowledge Graph entities. | The strongest tie to organic ranking of the six, but a loosening one, and studies disagree on how strong. Recent measurements of top-10 overlap range widely, and a large share of citations now come from pages ranking 11 to 100 or from formats like YouTube. Ranking in at least the top 20 still functions as a practical prerequisite. | Strong organic ranking plus Knowledge Graph presence are prerequisites. Schema markup and entity-linking are the technical levers. |
| Google AI Mode | Google Search’s conversational AI experience. Uses query fan-out, breaking one question into many sub-queries, which pulls more sources into a single answer, including content that wouldn’t rank on a first-page SERP. | Loose dependence by design. Fan-out reaches well past page one, and analyses find the large majority of its citations come from outside the organic top 10. Depth across a cluster of related pages predicts visibility better than a single top ranking. | Cluster content covering adjacent questions performs well. Single-page authority is less important than topical depth across a site. |
| Perplexity | Search-first AI built around citation-heavy answers, typically eight or more sources per response. Heavy use of Reddit, YouTube, news, primary research. Strongest recency weighting of the major engines. Most likely to refuse rather than guess. | Among the higher dependences on organic rank, though it varies sharply by industry. It leans on authoritative, already-ranking domains and journalism sources, so a strong organic footprint and earned media both feed Perplexity citations. | Authoritative on-domain content unlocks Perplexity answers. The more first-party material exists, the more Perplexity engages. |
There is a pattern worth naming. Engines that are more willing to generate confident answers (Gemini and AI Mode, in particular) produce higher visibility but carry a higher risk of inaccuracy. Engines that are more willing to hedge or refuse (Claude and Perplexity) produce lower visibility but higher credibility. There is a direct trade-off between confidence and accuracy at the engine level, and your business is being represented by all six simultaneously.
One engine, two systems: what changes when ChatGPT thinks harder
The table above treats each engine as one system. For ChatGPT, that is not quite true, and the gap is larger than the gap between some of the engines in the table.
ChatGPT answers in two modes. Instant mode is the fast default. Thinking mode does deeper multi-step research. Semrush, working with Kevin Indig, ran 100 prompts through the model twice, once in each mode, and compared what came back¹³.
Only 25.6% of the cited domains were the same. Roughly three in four sources changed for the same question, on the same engine, on the same day.
WHAT CHANGES WHEN CHATGPT THINKS HARDER
Instant mode Thinking mode
User-generated and review sites
Official documentation and support pages
Government and academic sources
Only 25.6% of cited domains are the same across the two modes.
Share of all citations, from 100 prompts run twice through the same model. Bar lengths are proportional to each other throughout this chart.
The rest of the behaviour changes with it. The share of answers citing any external source rises from 50% to 68%. Sources per answer nearly double, from 2.6 to 4.5. The model runs 4.6 times as many internal sub-searches before answering, and across the full test it ran 1,130 web searches in Thinking mode against 245 in Instant. It reached 173 unique domains rather than 127, and 99 of those never appeared in Instant mode.
The source mix is what a business owner should look at
The two modes cite differently, not merely more often, and the direction is consistent.
Reddit falls from 15% of citations to 7%. Other user-generated and review content falls from 14.3% to 6%. Meanwhile government and academic sources roughly quadruple, from 1.9% to 8.8%, and official documentation and support pages grow from 12.4% to 17.5%. Brand-owned sources hold roughly steady at around 61%.
So the softer social proof that carries a business in Instant mode loses about half its weight in Thinking mode, and formal, documented, attributable material takes its place. Indig’s summary is that the brand winning under one mode is not the brand winning under the other, and that these are two different systems.
This reaches you even if you never touch the setting
The obvious response is that most buyers never switch modes. That is true and it does not help, because ChatGPT routes complex prompts into higher reasoning automatically, including for people on the free tier. Comparisons, evaluations, and multi-criteria decisions are exactly the prompts that get routed, and supplier selection is exactly that shape.
The study also found the gap is widest at the start of the buyer’s journey. On early “how do I know if I need this” questions, the citation rate in Thinking mode ran 35 percentage points higher than in Instant. By the late validation stage the gap had narrowed to 5 points. Early-stage content is where the difference is decided, which is an argument against treating it as brand awareness with no measurable return.
One more finding worth knowing: under Thinking mode, a brand cited at the start of a conversation often stayed cited through to the end, in four of the twenty journeys tested. Under Instant mode that happened in none of them.
What this means for the work
It does not change what you should publish. It sharpens why.
Documentation, specifications, methodology pages, dated case studies and original data are the material Thinking mode reaches for. Reviews, forum presence and third-party mentions are what Instant mode leans on. Most businesses have invested in one and not the other, and until now there was no clear reason to do both.
**One caveat, in keeping with the rest of this page.** This is a single study of 100 prompts on one model version, published in mid-2026. The size of these effects will move as the models change. The mechanism behind it, that deeper research reaches different and more formal sources, is the part worth planning around.
How much your Google ranking still matters
The third column in that table raises an obvious question: if AI engines pull from different indexes and weight ranking so differently, is traditional SEO still worth doing? The short answer is yes, but its job has changed.
Two facts have to sit side by side. The first is that organic ranking still runs most of search. On the many queries where Google shows no AI Overview, the organic results are the entire experience, and a customer who clicks a blue link is choosing among ranked pages the old way. The second is that for the engines and queries where AI does answer, the link between a top ranking and a citation has loosened, and it varies by engine. The studies that measure how often AI-cited pages also rank in Google’s top 10 disagree by a wide margin. Some recent ones put the overlap near the high end, others near the low end. They agree on the direction even when they disagree on the size: ranking well still helps, but it is no longer enough on its own.
Reading across the engines, three groups emerge. Google AI Overviews and Perplexity lean most heavily on organic authority; strong rankings and authoritative, already-ranking domains feed their citations. Gemini and Google AI Mode sit in the middle, where rank is a gate (you generally need to be in contention, often the top 20) but clarity, structure, and topical depth decide the cite. ChatGPT and Claude lean least on your Google position, because they retrieve through Bing and Brave and weight brand mentions and content depth over where you rank on Google.
That spread is why foundational SEO remains the floor under AI visibility, not a separate track. A site that isn’t crawlable, fast, structured, and organized around real topics won’t rank, won’t be indexed cleanly by any of the engines’ source indexes, and won’t provide engines with clean passages to extract. The technical and structural work that earns a Google ranking is the same work that makes a page citable. What’s changed is the goal you optimize toward on top of that floor: not the single keyword and the single ranking position, but topical depth across a cluster of related questions, clear passage-level answers an engine can lift, structured data that disambiguates your business as an entity, and freshness signals that let current content displace stale content.
So the SEO program doesn’t go away. It becomes the foundation, and the AI Visibility work is built on top of it. Our Foundational SEO page covers the technical and on-page groundwork that all of these engines rely on before any AI-specific optimization can work. That includes crawlability, site structure, schema, and content organization.
Is this getting better? What the labs are doing about it
A reasonable next question is whether AI engines are getting better at distinguishing source material from pattern-matching, and whether the developers behind them are working to make this less of a problem over time. The answer is mixed.
What’s improving. The major labs publish steady evidence that hallucination rates are declining across model generations. OpenAI’s GPT-5 system card reports that hallucinations in challenging conversations are reduced by 8x compared with the prior generation, and that GPT-5 responses with web search enabled are approximately 45% less likely to contain a factual error than GPT-4o¹. Google DeepMind has published the FACTS Grounding benchmark, a public methodology for measuring how well models ground their answers in provided source material², now extended into a four-part suite covering grounding, search, multimodal, and parametric factuality³. Several techniques are moving from research into production: retrieval-augmented generation (RAG), which forces the model to draw from a specific retrieved corpus rather than generate freely; uncertainty-aware training methods that reward abstention when evidence is thin; and span-level verification that flags claims the system cannot support before they reach the user.
What’s structurally hard. In September 2025, OpenAI published research explicitly identifying why this is difficult to fix at the model level. The training and evaluation procedures that produced these systems reward guessing over abstention. A model that says “I don’t know” scores worse on accuracy benchmarks than a model that guesses and is occasionally right⁴. Earlier academic work formally argued that hallucination cannot be entirely eliminated from systems built this way; it can only be reduced⁵. Practitioners and labs largely converge on this assessment: hallucinations will decline. They will not disappear.
What’s changing about freshness. Engines have begun weighting recency signals more heavily. Gemini specifically has integrated last-modified dates and Article schema freshness into its source selection. ChatGPT’s live web search now factors in publication recency. Perplexity has always been recency-biased. That’s structural to its design. These changes don’t solve the stale-data problem, but they make it possible to displace stale source material more effectively if you publish current content with clean freshness signals.
What this means strategically. Treat AI engine behavior the way you treat Google algorithm changes. The engines update continuously. The specifics of what they do well and what they fail at will keep shifting. The underlying mechanics are stable enough to plan around. These are retrieval versus generation, the role of source material, and the freshness premium. The tactical details (which engine weights what signal how heavily) are not. This is why ongoing visibility work matters more than one-time audits. The first ninety days of an engagement establish the foundation; the work after that adjusts as the engines evolve.
Why the same question can produce a different answer next week
The answers move. If you ask an engine the same question today and again next Tuesday, you will often get different wording, different businesses named, and different sources cited. This unsettles people the first time they see it, and it is worth understanding properly, because the conclusion most people jump to is wrong.
Several independent studies put numbers on it. Ahrefs analyzed more than 43,000 keywords and found that an AI Overview has roughly a 70% chance of changing between one observation and the next, with the content lasting about two days on average⁶. Authoritas measured AI Overview volatility at 0.68 over an eight-week window against 0.49 for standard organic results, so the AI layer moves faster than the blue links underneath it⁷. Semrush, studying a month of consistently appearing AI Overviews back in late 2024, found individual pages holding their citation for an average of just under four consecutive days, with 91% of tracked pages dropping out at some point and fewer than half ever returning⁸.
Then comes the part that usually gets left out. Ahrefs also measured whether the meaning changed, and found consecutive answers scoring 0.95 out of 1.0 for semantic similarity⁶. The engines are rewording a stable conclusion, not changing their mind. Ask two experts the same question and you get different sentences and the same answer. That is what is happening here.
THREE LAYERS, THREE SPEEDS
The wordingMOVES FASTEST
The engine rephrases the same conclusion constantly. Ahrefs found consecutive answers scoring 0.95 out of 1.0 for semantic similarity. This layer means the least.
The businesses namedMOVES SLOWLY
Who the engine actually recommends. This is the layer that decides whether a buyer sees you, and it is the one worth watching.
The citationsROTATE HARDEST
Individual pages hold their place for a few days at a time, and they rotate whether or not anything on your site changes.
Reading all three as one number is what produces the panic.
Reading those three layers as one number is what produces the panic. A business that watches its wording change daily concludes it is invisible. A business that watches the named list is looking at something real.
Volatility tracks retrieval
The pattern follows the retrieval and generation distinction this page opened with. Engines that search the web at the moment you ask inherit the movement of the results they find, plus the natural variation of generating language. Engines that lean more on what the model already holds drift more slowly. That is why the same brand can look stable on one engine and restless on another in the same week, and why a single engine’s answer is a poor proxy for the whole picture.
Specific questions are steadier than vague ones
This is the finding that matters most for a local or regional business, and it is good news.
SE Ranking tested 5,000 local queries in Google AI Mode and found that phrasing drives volatility more than geography does. Vague questions of the “near me” kind returned only about 35% of the same websites when repeated in the same city. Questions that named the place, such as “restaurants in Denver,” nearly doubled that consistency, to roughly half the same websites and pages holding across runs. In more than a third of those location-specific searches, the overlap ran above 90%⁹.
A buyer looking for a Connecticut supplier usually names Connecticut. Those questions sit in the steadier half of the range.
What this means for how you measure
One answer is an observation, not a verdict. Read one and you learn what one engine said on one morning to one person in one place. Read a pattern across many answers and you learn something durable about how the engines see your category and where you sit in it.
That is why the diagnostics here read patterns rather than moments, and why a free snapshot is designed as a triage instrument, a way to find out whether there is something worth investigating, rather than a measurement of your standing. It is also why the deeper engagements run across weeks: an accumulated record survives the daily movement in a way that a single check never can.
The practical version, for an owner: do not celebrate one good answer, and do not lose a week over one bad one. Ask whether the engines consistently connect your business to the category you sell in. Look at what they cite when they answer questions in your field, and get your business into those sources. Companies that appear in most answers most of the time earned it by being well documented across many places, not by winning one query.
Two recent shifts and what they reveal about how AI engines answer questions
Two changes in the last twelve months illustrate how AI search behaves in practice, and how the work of being visible needs to keep up.
The Google Business Profile Q&A deprecation
In late 2025, Google discontinued the user-managed Q&A feature on Google Business Profiles. The API that let third-party tools post and manage Q&A shut down on November 3, 2025, and the public-facing section began rolling off listings on December 3, 2025. For roughly a decade, business owners and customers could post and answer questions directly on a business’s profile. Those questions and answers showed up in Google Search and on Maps, and AI engines reading the profile would pull from them.
Google replaced the feature with Ask Maps, a Gemini-powered feature that generates answers about businesses in real time. Ask Maps reads the GBP fields, the reviews, the photos, and the website content, then produces a response. The answer is not a user-curated Q&A; it is an AI-generated response based on whatever source material Google can pull at the moment of the question.
The implication for business owners is direct. Before the deprecation, an owner could draft a question-and-answer pair to clarify something specific about the business (service area, payment options, after-hours policies) and post it for AI engines to read. After the deprecation, the owner cannot directly seed answers. The owner can only shape the source material Ask Maps reads. That means the GBP fields, the reviews, the photos, and especially the company’s website.
The mechanics from the top of this page apply directly. Ask Maps is doing retrieval-plus-generation in the moment a customer asks a question. When the source material is rich and specific, Ask Maps retrieves and the answer is grounded. Rich and specific means a current GBP listing with a complete description, recent photos, current staff, an updated services list, plus website content with structured FAQ blocks on the cornerstone topics. When the source material is thin, the generation portion grows. Owners cannot post the answer they want; they can only put the right material where Ask Maps will find it.
The practical replacement for the deprecated Q&A field at the website level is FAQ blocks on the company’s cornerstone landing pages, with FAQPage schema markup. These give Ask Maps and the other AI engines the structured question-and-answer material they previously pulled from the GBP Q&A field. The format is the same; the location has changed.
The asymmetric reading of credentials
A second pattern, observed across recent engagements: different AI engines can read the same fact about a business differently, and the way they read it does not always match common sense.
The illustrative case is industry-association membership. Take a business with a strong operational record. It is a fifty-year-old, family-owned service company with zero complaints lodged with the Better Business Bureau over its entire history. That business has chosen not to participate in BBB’s accreditation program; the founders’ position was that the substantive record is the relevant evidence rather than the accreditation status.
How do AI engines read that? They split.
Some engines (Claude in particular, in recent observations) weigh the substantive record of fifty years and zero complaints as structural evidence of operational quality. The engine reads the substance and presents the business favorably.
Other engines (ChatGPT, Perplexity, Gemini in the same observations) read the positional signal of “not accredited or rated” reductively. The engine sees the absence of an accreditation badge and flags it as a credibility concern, even though the substantive record is stronger than most accredited businesses can claim.
The split happens because the substance and the positional signal exist in different parts of the source corpus the engines read. The accreditation status appears in structured data such as the BBB profile’s headline status, directory listings, and the Knowledge Graph. The fifty-year record appears in less-structured forms such as the company’s About page, customer reviews, and news coverage. Engines that weigh structured signals heavily pick up the positional reading; engines that weigh narrative reasoning and structural commitments pick up the substantive reading.
The implication is that businesses with strong operational records but no industry-association memberships often have an asymmetric reading problem they don’t realize they have. Joining the association the company has principally declined would only concede the positional reading. The better move is to make the substantive record visible, in the GBP description, on the About page, on the cornerstone landing pages, in review responses where it fits, and in structured data where the schema supports it. The substance becomes the citation rather than the status.
This is the deeper pattern that connects both case studies. AI engines read what is in front of them. The work of being visible is making sure the right material is in front of them.
Ads have arrived inside the answers, and they do not buy you the answer
The third development is the newest, and it changes what a customer sees when they ask about your category.
These answers used to be unmonetized. That ended quickly. Google began placing ads inside AI Mode responses in late 2025, and SE Ranking’s study of 50,032 commercial keywords, collected in June 2026, found a text ad on 29.45% of them, close to one query in three¹¹. In 71.1% of those cases, two competing advertisers appeared in the same block. Cost per click predicted placement better than anything else: ad presence ran at 24.33% on keywords under $2 and 53.56% on keywords at $10 and above. It varies enormously by industry, from 72% in pets down to under 3% in healthcare. ChatGPT added ads on a similar timeline, as a labeled sponsored card below the answer rather than inside it, shown on its free tiers. We have captured both formats in our own monthly research, sitting directly beneath answers to ordinary buyer questions.
Paying for the slot does not buy the citation
Here is the finding that matters, and it is the reason this section exists.
SE Ranking checked whether the advertisers appearing in AI Mode were also among the sources the engine cited for that same query. They almost never were. Only 11.53% of advertiser domains appeared in the citation list, and at the exact page level, 1.95%. Put the other way around, for 88% of the keywords carrying an ad, the advertiser was not among the sources the answer was built from¹¹.
The obvious objection is that advertisers might simply be weaker sites. SE Ranking tested that, comparing advertisers against non-advertising domains of equivalent strength for the same query, matched on domain trust, backlinks, referring domains, and organic standing. The advertisers were cited no more often. Separately, for about 85% of the keywords they advertised on, those advertisers did not appear anywhere in the organic results, on any page.
So three things are running side by side on the same result, and each is earned separately. The ad slot is bought. The organic ranking is earned through the technical and structural work. The citation, which is the engine naming you inside its answer, is earned through authority and content. Money moves the first one. It does not move the third.
THREE CHANNELS, EARNED THREE DIFFERENT WAYS
THE AD SLOT
A labelled advertisement beside or below the answer. Present on 29.45% of commercial keywords in AI Mode.BOUGHT
THE ORGANIC RANKING
Your position in the results underneath. Earned through technical and structural work on your site.EARNED
THE CITATION
The engine naming you inside its answer. Only 11.53% of advertisers were also cited for the same query.EARNED, NOT FOR SALE
Money moves the first one. It does not move the third.
Being cited does lift the ad: advertisers cited in the answer ran roughly four to seven percentage points higher on paid click-through than those who were not.
So what does the ad buy?
A fair question follows: if the ad sits at the bottom and does not get you into the answer, why buy it? It turns out there is a good reason, and the numbers are more interesting than the ones above.
On Google, paid is the one thing on these results that has held up. Seer Interactive’s full-year study found paid click-through rates on AI Overview results staying in a 13% to 16% band through 2025 and into 2026, while organic click-through on those same results fell to a floor of about 1.3% before recovering slightly¹⁰. Averaged across the year on informational queries, a brand cited in the AI Overview earned about 2.1% organic click-through and about 15.7% paid. So on a result where an AI answer has absorbed most of the clicks, the ad is often the most reliable way left to get one. That is a real answer to why advertisers are buying, and it is why paid budgets have moved toward these results rather than away from them.
Then comes the part worth sitting with. Being cited in the answer does not just help your organic performance; it makes your ad perform better too. In the same dataset, advertisers whose brand was cited in the AI Overview ran roughly four to seven percentage points higher on paid click-through than advertisers who were not cited, in nearly every month measured¹⁰. The reader appears to see the brand inside the answer, then sees the same name on the ad, and clicks at a materially higher rate.
The two are not substitutes. They compound. Content and authority work earns the citation, the citation lifts the ad, and the ad captures a click the organic result can no longer reliably win.
ChatGPT is a different and much less settled picture, and it deserves a plainer caveat. Its ads are sold by impressions rather than clicks, at a reported premium to search advertising, and advertisers receive aggregate reporting rather than the measurement they are used to. Trade analysts have read the pricing model itself as a signal that the format is being sold as exposure rather than as direct response¹². Whether those placements drive traffic at a rate that justifies the cost is not yet publicly established, and anyone telling you otherwise is ahead of the evidence.
What this means for a business owner
Being the cited source is worth more than it was a year ago, not less. It is the one position that cannot be bought, it is what a reader trusts, and on the paid side it now measurably improves the performance of the advertising you are already running.
An ad is a reasonable thing to buy on these results. It is not a substitute for the work on this page. If you buy the placement while your business is absent from the answer above it, you are renting attention next to a recommendation of someone else, and paying the full rate for a click that a cited competitor gets more cheaply. Whether that trade is worth making is a paid-media decision with its own arithmetic, and for some businesses it clearly is. What it is not is an AI visibility strategy, and it should not be sold to you as one.
What this means for the work ahead
Three principles follow from how these systems behave. They shape every recommendation we make in an AI Visibility engagement.
First, current authoritative content displaces stale content. AI engines pull from whatever is indexed and accessible. The 2015 review that mentions a staff member who no longer works there is still being read. The blog post from 2018 with outdated information is still being treated as current. The way to fix that is not to delete the old content. It is to publish new content that is more specific, more recent, and more relevant, so AI engines have something better to pull from.
Second, source diversity reduces risk. For many businesses, a single review platform or a small set of citations carries most of the weight in shaping what AI engines say about them. If that platform changes, gets purchased, or goes offline, the source material disappears with it. Diversifying across Google Business Profile, Yelp, BBB, industry-specific platforms, and the company’s own domain spreads the risk and strengthens the picture.
Third, the company’s own website is the foundation. Everything else points to it. Reviews mention it. Citations link to it. AI engines pull from it directly. When the website has thin or outdated content about a topic AI engines get asked about, the engine fills the gap with generated material. When the website has rich, current, well-structured content on that topic, the engine has something to retrieve. This is why the cornerstone landing pages we recommend matter so much. They are the source material AI engines will pull from for the next several years, on the questions your customers ask most often.
The process isn’t complicated. It is sustained. Every new piece of authoritative content, every additional review, every clean citation source adds to the pool AI engines pull from. The cumulative effect is what changes how AI describes your business.
For most businesses, that comes down to six things:
SIX KEY THINGS YOU CAN DO RIGHT NOW
Publish current, specific content
Give engines something better to retrieve than old material.
Keep your Google Business Profile current
Description, services, hours, photos, and staff.
Check that engines can find you on a map
Your listing and your website are separate records. Ask more than one engine to show your business on a map, and confirm the category is right and the listing does not say permanently closed.
Add FAQ blocks with FAQPage schema
On cornerstone pages, where the old GBP Q&A now lives.
Diversify your review and citation sources
Spread the weight across several platforms.
Give every page clean freshness signals
So current content can displace stale content.
The work isn’t complicated. It’s sustained.
Frequently asked questions
How do AI engines build answers about my business?
They use two paths: retrieval and generation. Retrieval pulls from real source material such as your website, reviews, business listings, and citations. The engine cites what it used. Generation predicts plausible text when source material is thin. Most answers mix both, and the less source material exists, the larger the generated portion grows.
What do AI engines pull from?
Five main places: training data fixed at the model’s cutoff, live web search at the moment of the question, the specific pages an engine cites, business directories and review platforms, and Google’s Knowledge Graph. Engines do not reliably tell current content from old content, so stale material and current material can carry equal weight.
Why does AI visibility take different work than SEO?
Traditional SEO aimed at ranking a link the reader still had to click. AI search reads the sources, decides what to say, and presents the summary as the answer. Most people never click through. The work shifts from earning a rank to getting accurate, current source material into the places engines read. Foundational SEO is still the floor it all sits on, though: a site that can’t be crawled, structured, and indexed cleanly won’t be cited by any engine.
Does my Google ranking still matter for AI search?
Yes, but how much depends on the engine. Google AI Overviews and Perplexity lean most on organic authority. Gemini and Google AI Mode treat rank as a gate but let clarity and depth decide the citation. ChatGPT and Claude lean least on your Google position because they retrieve through Bing and Brave. Across all of them, ranking well still helps but is no longer sufficient on its own.
Are AI engines getting better at avoiding made-up answers?
Partly. Hallucination rates are falling across model generations, and the labs are publishing steady evidence of this. But research from the labs themselves shows the problem can’t be eliminated entirely in systems built this way. It can be reduced, not removed.
What does my business need to do to be represented accurately?
Publish up-to-date, specific content about the questions your customers ask, so search engines have better material to retrieve than old content. Keep your Google Business Profile complete and up to date. That means the description, services, hours, photos, and staff. Add FAQ blocks with FAQPage schema markup on your cornerstone pages, since that is now where the old GBP Q&A material lives. Diversify your review and citation sources rather than relying on a single platform. Give every page clean, fresh signals, so current content can displace stale content.
If AI answers keep changing, is any of this worth measuring?
Yes, and the change is part of what makes it worth measuring. Constant movement means the shortlist is being reconsidered all the time, so a well-documented business gets repeated chances to enter it rather than facing a fixed order it has to displace. What changes is how you measure: aggregate patterns across many answers rather than single checks. Ahrefs, having measured the volatility, reached the same conclusion, recommending that businesses track visibility across a volume of answers and aim to be associated with a topic rather than to win one query.
My business appeared in an AI answer last month and it is gone now. What happened?
Usually nothing you did. Cited pages rotate constantly, and a business can leave and re-enter an answer set without anything changing on its website. The question worth asking is not whether you appeared once, but whether you appear consistently. If you are named in most answers to the questions your buyers ask, you have durable visibility. If you appeared once and have not since, that first appearance was more likely the natural movement of the answers than a position you held and lost.
Can I pay to appear in AI answers?
You can pay to appear beside them, not inside them. Research on 50,032 commercial keywords found that advertisers appeared among the sources the engine cited only 11.53% of the time, and the advantage disappeared entirely once advertisers were compared against equally strong sites that did not advertise. The ad slot and the citation are separate things, earned separately.
That does not make the ad pointless. On Google results carrying an AI Overview, paid click-through has held up far better than organic, so the ad is often the most reliable remaining way to earn a click. But the two work together rather than instead of each other: advertisers who were also cited in the answer ran several percentage points higher on paid click-through than advertisers who were not. Earning the citation makes the advertising you are already buying work harder.
References
- OpenAI. (2025, August 13). GPT-5 system card. OpenAI. https://openai.com/index/gpt-5-system-card/
- Google DeepMind. (2024, December 17). FACTS Grounding: A new benchmark for evaluating the factuality of large language models. Google DeepMind Blog. https://deepmind.google/blog/facts-grounding-a-new-benchmark-for-evaluating-the-factuality-of-large-language-models/
- Google DeepMind. (2026, March 3). FACTS Benchmark Suite: a new way to systematically evaluate LLMs’ factuality. Google DeepMind Blog. https://deepmind.google/blog/facts-benchmark-suite-systematically-evaluating-the-factuality-of-large-language-models/
- Kalai, A. T., Nachum, O., Vempala, S. S., & Zhang, E. (2025, September 4). Why language models hallucinate (arXiv:2509.04664). arXiv. https://arxiv.org/abs/2509.04664
- Xu, Z., Jain, S., & Kankanhalli, M. (2024, January 22). Hallucination is inevitable: An innate limitation of large language models (arXiv:2401.11817). arXiv. https://arxiv.org/abs/2401.11817
- Linehan, L., & Guan, X. (2025, November 11). AI Overviews change every 2 days (but never change their mind). Ahrefs Blog. https://ahrefs.com/blog/ai-overview-change/
- Authoritas. (2025). SERP organic and AI Overview volatility research. Authoritas. https://www.authoritas.com/blog/serp-organic-and-ai-overview-volatility-research
- Oberstein, M. (2024, November 20). Exploring URL volatility in Google’s AI Overviews. Semrush Blog. https://www.semrush.com/blog/url-volatility-ai-overviews/
- Deda, Y. (2025, September 29). AI Mode volatility: Results from our local search test. SE Ranking Blog. https://seranking.com/blog/ai-mode-volatility-test/
- McDonald, T., Cooley, H., & Williams, M. (2026, April 24). AIO impact on Google CTR: 2026 update. Seer Interactive. https://www.seerinteractive.com/insights/aio-impact-on-google-ctr-2026-update
- Khromova, Y. (2026, July 14). AI Mode now shows ads on nearly 1 in 3 commercial queries. SE Ranking Blog. https://seranking.com/blog/google-ai-mode-ads/
- EMARKETER. (2026). OpenAI’s ChatGPT ads reshape digital playbook with view-based pricing. EMARKETER. https://www.emarketer.com/content/openai-chatgpt-ads-reshape-digital-playbook-with-view-based-pricing
- Loktionova, M. (2026, June 30). *Only 25% of cited sources overlap between ChatGPT’s different reasoning modes [Study]*. Semrush Blog, in partnership with Kevin Indig. https://www.semrush.com/blog/chatgpt-reasoning-ai-visibility/
Curious about how your business shows up in AI engine answers?
If you want the data first, start with a free AI Visibility Snapshot. If you’d rather talk it through, schedule a 30-minute conversation.
Get Your Visibility Snapshot → Schedule a Conversation →The CT Brief is a free monthly look at how AI engines answer real questions about Connecticut businesses.
Get the CT Brief →About the Practice
Telstar Consulting is an independent AI Visibility practice based in Connecticut. It helps businesses show up accurately and often when buyers ask AI engines for recommendations, with SEO as the foundation. That matters because more buyers now begin inside AI engines than on search results pages, and a business the engines don’t name doesn’t make the shortlist.
Explore the services →