Different AI, different internet

Ask two AI assistants the same question and you get two answers. Often they rest on different sources as well. We wrote earlier that models differ in their values (Every AI model has an opinion) and in their skills (our exam report card). Those differences sit at the end of the chain. The divergence starts sooner, while the assistant is still gathering its material.

Between your question and the answer sit four steps. At each one assistants make different choices: the search terms they use, the index they search, whether they read the pages or only the excerpts the search engine returns, and how they interpret what they read. This piece walks through those steps, and then looks at what they mean for being found and for the quality of the answers you rely on.

Four steps from question to answer

Step 1: from question to search terms

An assistant does not search for your question as you typed it. It first turns it into search terms of its own: sometimes one query, sometimes a dozen. Google describes this for its AI Mode as query fan-out, splitting a question into subtopics and searching each one separately. Which words the model chooses, in which language, and which aspects it treats as part of the question: all judgement calls. Two assistants make them differently, and no later step can recover a source the search terms never reached.

Your own framing travels along. A question that already contains its answer, such as “why is option A better than option B?”, tends to produce search terms that look for confirmation. Most interfaces do not show which queries the assistant actually ran.

Step 2: from search terms to results

The search terms then go to a search index, and each of the large assistants searches a different one. Each index has crawled its own part of the web and ranks it in its own way, and the model usually looks only at the top handful of results.

The overlap is small. Writesonic ran 161,286 prompts through ChatGPT, Gemini, Perplexity and Google AI Overviews in May and June 2026. On the same prompt, any two of them shared on average about 17 percent of the websites they cited. Roughly 72 percent of all cited websites appeared in one engine only. ChatGPT, which leans on Bing, overlaps least with the other three.

Step 3: reading the page or only the snippet

This step gets the least attention. A search result is a title, a link and a short excerpt, picked by the search engine because it matches the search terms. The assistant can answer from those excerpts, or open the pages and read them. AI providers treat these as two separate actions in their own tools: searching and fetching. Fetching costs time and tokens.

An answer built on snippets is second-hand information. The excerpt was selected to match words, not to represent the page, and it drops what you need to judge a claim: the date, the conditions, the sentence before it that says “contrary to what is often claimed”, or who is speaking. An assistant that opens the page can check a claim in its context. One that reads only snippets repeats what the search engine chose to show. Where speed counts, the balance tips towards snippets. In our experience Claude's voice mode, for example, answers from the snippets without opening the pages.

Sometimes there is no choice. A page that blocks AI fetchers, sits behind bot protection or only renders through JavaScript leaves the snippet as the only text to read. Under the answer, the citation looks the same in every case. A source link shows where information came from, not whether the model read the page.

Step 4: from reading to answer

Only now does the part we wrote about before come in. The model decides which sources to believe: research we discussed in Every AI model has an opinion found that models prefer institutional sources and can be swayed by a claim that simply appears more often. Then it has to read well. In our exam benchmark, Dutch was the subject the models struggled with most, because that exam tests reading: what does the author actually believe, and which argument is left unspoken. Models turned out to be intelligent in different ways: excellent in one subject, failing in another.

Finally the model writes one fluent answer. When the sources contradict each other, that conflict often disappears in the wording, and the answer sounds as certain as one built on agreement.

Each step narrows what the next one gets to work with. Two assistants can part ways at all four, and the answer alone does not show where.

What this means for being found

For twenty years, being found online meant ranking on Google. With AI assistants the question becomes whether an assistant reaches your page, reads it and represents it correctly. Each of the four steps adds its own condition.

Answer the question behind the search terms

Assistants search with their own reformulations, often several per question. Pages that answer concrete questions in plain language score better across those varied queries, because they match the many ways people phrase a question rather than targeting a single keyword. A page that also covers the sub-questions of a topic gets more chances when a question fans out.

Be visible in more than one index

A top position on Google says little about ChatGPT or Perplexity. Ahrefs found that of the pages AI assistants cited, only about 12 percent ranked in Google's top ten for the same question. Search for your organisation and key topics on bing.com and search.brave.com as well. Missing pages can be submitted through Bing Webmaster Tools, and Bing supports IndexNow for announcing new ones.

Write for the snippet, and keep the page readable

Assume that sometimes only the excerpt gets read. The opening sentences of a page, and of each section, should carry the core facts on their own: what, for whom, how much, as of when. A claim that only holds after three paragraphs of context will be quoted without that context.

Then check that an assistant can open the page at all. AI companies use separate bots for training, for building a search index and for fetching a page when a user asks for it. Each has its own name in robots.txt, so you can refuse training without disappearing from the answers.

Bot protection matters too: Cloudflare has blocked AI crawlers by default since July 2025, unless the site owner lets them in. Content that only appears through JavaScript is another blind spot. A study by Vercel found that the crawlers of OpenAI and Anthropic download JavaScript files but do not run them.

Say the same thing everywhere

In the last step the model weighs sources against each other, and repetition counts. If your website, a comparison site and an old press release state different prices or features, the model picks one, and not necessarily yours. Keeping your facts consistent across your own pages, and getting outdated third-party pages corrected, does more than a new keyword. Measure per assistant too: ask the questions your customers ask. Note who gets cited and repeat that every few months.

What this means for the quality of answers

Your question steers the search

Because the assistant builds its search terms from your words, a leading question tends to return leading sources. Ask neutrally, and ask explicitly for evidence against the conclusion you expect. If the interface shows which searches were run, look at them: that is the question that was actually researched.

What was not found stays invisible

The source list under an answer shows what was found, never what was missing. An answer resting on five pages from one index sounds just as confident as one resting on fifty from several indexes. For a quick fact that rarely matters. For a policy question or the choice of a supplier, the missing part can be the part that mattered.

First-hand or second-hand

For claims you are going to act on, find out whether the assistant read the source or only its excerpt. Ask it to open the page and quote the relevant passage, or open it yourself. Be extra careful with quick and spoken answers, where the pressure to skip the reading is highest.

Agreement counts when the routes differ

Journalists confirm a story with two independent sources. With AI, independence means the routes differ: other search terms, another index, pages actually read, another model doing the interpreting. Two assistants that agree after different routes give you more confidence than two that searched the same index and skimmed the same snippets. Do assistants on different routes contradict each other? Pay attention: the evidence is often thinner or more contested than either answer suggests.

How Localign uses these differences

Most questions do not need a broad search. For everyday work, Localign picks the smallest model that is still accurate enough for the task. That saves tokens, and with them energy.

For contested or sensitive topics, or whenever you ask for it, Localign runs several agents side by side. Each with a different model and a different search engine. Their routes differ at every step described above, so their results complement each other. Say you are researching climate policy, or comparing options for a new car: an agent searching through Perplexity and one searching through Claude return different source lists and different readings. Localign merges those sources and interpretations into one summary.

Admins decide which external providers are switched on. What goes out to a search engine is the underlying question. Identifying details from your conversation stay inside Localign, unless you give permission for a specific detail the search genuinely needs.

Compare models yourself in Localign

Localign normally picks the model that best fits your question. You can also name a model yourself in your prompt, such as ChatGPT or Claude. If you want to use several models, you say so in your prompt as well. Each model then gives its own answer, and the replies are merged into one. That way you combine different perspectives and get the best out of several models. Under the answer you can see which models were used.

When you open a model, you can see the question Localign sent it, together with that AI model's full answer and the sources it consulted. That way you know which model answered what, and you can check the sources yourself instead of relying only on the summary.

With 'Compare models' you can put the answers side by side. For each model you see which question it received and which sources it used. In this example both models name Peppol and 1 January 2026. Claude says you must keep invoices for at least seven years, while ChatGPT points out that invoices to consumers fall outside the rule.

The sources differ too. ChatGPT cites one source: the government's e-invoicing site. None of the nine sources Claude used comes from a government website. That way you can compare the answers and the evidence behind them directly, side by side. This is the check from 'Agreement counts when the routes differ': two routes, compared on one screen.

The question Localign sends to a model is not literally what you typed. Localign rewrites your prompt without sending along privacy-sensitive information. That way your privacy-sensitive information stays inside Localign.