AI Search Engines: The Models Behind Each One, and the Indexes They Actually Read

Three layers behind every answer box

Ask most people what powers ChatGPT and they will say GPT. True, and not much use to a business. Every AI search product is three things stacked together. At the bottom sits an index: a crawled, ranked copy of the web, with map and product data beside it. In the middle sits a large language model (LLM) that reads what the index returns and writes the paragraph. On top sits the product, the app or search box a customer types into, with its own rules for links, locations and ads.

The layers move at different speeds, and only one of them is within reach. That is why this map is organized by where answers come from, not by whose model topped a benchmark this month. For how the layers combine into a shortlist of three names, read how AI chooses which businesses to recommend.

The LLMs behind the names, and who builds them

People search for a list of LLMs as if the model were the thing to optimize. A model on its own knows only what it absorbed in training, frozen at a cutoff date, and it never shows a link for that knowledge. It is the live search layer that names a business this week. Still, knowing who builds what helps, because one model family often sits inside several products.

Training has its own robots.txt switches, separate from search: GPTBot for OpenAI, ClaudeBot for Anthropic, Applebot-Extended for Apple and Google-Extended for Gemini. Blocking them keeps future content out of model training without removing you from search answers, with one wrinkle. Google-Extended also covers grounding in the Gemini apps, so blocking it reaches further than training. The AI crawler guide has every token in one table.

Four indexes feed most of what the engines say

Strip away the branding and the market runs on a short list of indexes. This is the useful map, because work done on one index pays off in every product that reads it.

Look at the overlap. A business that is indexed and ranked well in Google and Bing has covered the sources for five or six products. A business missing from Bing has a hole in two of the biggest, and may never find out, because almost nobody checks Bing. GEO vs SEO explains why ordinary rankings carry so much of the load.

Eight products compared on the points you can act on

Each engine page goes deep on its own machinery. This table compares them at a glance on the two things an owner can actually do something about: which crawler has to be let in, and whether the company gives site owners any report at all.

Two columns deserve a second look. The crawler column is where most self-inflicted damage happens: a firewall rule written to stop 'AI scrapers' that also stops OAI-SearchBot or PerplexityBot takes the site out of those answers. The reporting column explains why most owners measure AI search by anecdote. Only Google, Microsoft and Apple hand site owners first-party numbers, and none of them shows the answer a customer actually read.

Local questions run on borrowed listings

Ask any of these engines for a plumber nearby and the web index is only half the story. Each one also leans on a listings source, and they do not all lean on the same one. This is how a single stale phone number gets repeated by three machines at once.

• Gemini uses public Google Maps information, including ratings, reviews and opening hours, so the Google Business Profile is its local record.
• AI Mode and AI Overviews draw on Business Profiles and, for stores, Merchant Center feeds, according to Google's own AI optimization guide.
• Perplexity has licensed Yelp data for local answers since 2024 and shows Tripadvisor details for hotels since 2025, according to press reports.
• Copilot sits on Bing, and Bing Places moved to bing.com/forbusiness in October 2025 with deeper Copilot ties planned.
• Siri and Apple Maps read the Apple Business place card; Apple Business replaced Apple Business Connect on April 14, 2026.
• ChatGPT drew 58% of its local sources from business websites in BrightLocal's November 2024 study, against 27% third-party mentions and 15% directories.

One business, six listing systems, six chances to disagree with yourself. Entity optimization is the cleanup, and the guide to fixing wrong business information in AI answers covers what to do once an engine has already repeated the error.

Where to spend your first ten hours

There is no prize for treating every engine equally. Audience and overlap set the order. Google reported more than 2.5 billion monthly users for AI Overviews in May 2026, and in July 2026 said AI Mode had passed 1 billion and the Gemini app had reached 950 million. Those three share one index and one listing. Bing comes next, because it feeds Copilot and, on the outside evidence, much of ChatGPT.

1. 1

   Get Google's records right

   Indexed pages with snippets allowed, a complete Business Profile, a Merchant Center feed if you sell products. One job, three products.

2. 2

   Get into Bing on purpose

   Verify the site in Bing Webmaster Tools, submit the sitemap, turn on IndexNow and make Bing Places match Google character for character.

3. 3

   Open the door to the independent crawlers

   Read robots.txt and any firewall or CDN bot rule for OAI-SearchBot, PerplexityBot, Claude-SearchBot and Applebot. Each is a separate yes or no.

4. 4

   Fix the borrowed listings

   Apple Business, Yelp and the directories your trade uses. Perplexity and Apple Maps read them whether or not Google does.

5. 5

   Ask the engines and keep the answers

   Run your customers' real questions through each product. Our free AI visibility check asks several engines live through their APIs and shows who was named; API answers can differ a little from the consumer apps.

The smaller engines, and how sure anyone can be

The big names are not the whole list. GA4's AI Assistant channel, added on May 13, 2026, names DeepSeek and Grok alongside ChatGPT, Gemini and Copilot. Meta runs Meta-WebIndexer to improve Meta AI search results. Amazon says content crawled by Amzn-SearchBot can appear in services like Alexa, and Alexa+ books through partners such as OpenTable and Vagaro. For most local businesses these deserve a robots.txt check and little more, for now.

Then there is folklore: that one model prefers a certain writing style, that a special file earns citations, that a vendor has seen inside the ranking. Google says no third-party tool has access to its internal AI systems. The llms.txt explainer tests the most popular of these beliefs against the evidence.

Watching seven engines without losing a week a month

Testing every product by hand works for about a month. Then it stops. The durable version is a fixed set of customer questions, asked of each engine on a schedule, with every answer kept. Prompt tracking covers choosing the questions, the GEO tool comparison shows which trackers cover which engines and what they cost, and tracking AI traffic in GA4 sorts out the referrers.

At DataCram, GEO is part of our monthly SEO programs, which start at $625 a month, with most local businesses landing between $625 and $2,250 per the SEO pricing page. We watch answers across these engines with AICiteBeam, the tracking platform we built, and fix the records behind them. A free GEO audit runs your questions through each engine on this map and returns a fix list with a fixed monthly quote. New to the subject? Start at the GEO hub.