SEO when customers ask ChatGPT: how a B2B company wins enquiries in 2026
In 2026 a B2B company wins enquiries from search the same way it did before, with one difference: the customer increasingly reads an AI summary before visiting any website. Google says plainly that optimising for generative AI search is still SEO, and a page has to be indexed and eligible for a snippet to appear in an AI Overview at all. The companies that win enquiries have pages that answer customers’ questions directly, publish their own data and tools, get mentioned beyond their own site, and measure enquiries rather than traffic.
In short
- Google treats optimising for AI answers as ordinary SEO: a page has to be indexed and eligible to be shown with a snippet.
- AI answers pick up pages that answer directly in the first paragraph and offer something of their own: data, a tool or first-hand experience.
- llms.txt is a hint, not a ranking factor. Google says its search engine ignores the file.
- Whether AI recommends you often depends on what happens off your site: Business Profile reviews, directories and genuine mentions.
- Measure enquiries by source, not just traffic. More visibility doesn’t yet mean more customers.
I’m writing about what I do in my own businesses. I run SEO for reefa.pl, the site of my cleaning company in Kraków and Katowice, and for the KtoSprząta.pl directory. I give the numbers from those projects with dates. Some of them are baselines, meaning the state before the work, and I label them that way.
What changed in search
A customer looking for a supplier now has three routes instead of one. They type a question into Google and often see an AI Overview at the top, with links to its sources. They ask ChatGPT, which cites sources when it searches the web. Or they use Perplexity, which was built as a search engine that answers. Each of them summarises a handful of pages and links to only some of them.
The operators themselves publish some of the rules:
- Google builds AI Overviews on its own index and may run several related searches at once (query fan-out) to gather material for an answer. Beyond being indexed and eligible for a snippet, there are no extra technical requirements (Google’s documentation on AI features in Search).
- ChatGPT finds pages for its answers with a crawler called OAI-SearchBot. Sites that block it won’t be shown in ChatGPT search answers. A separate crawler, GPTBot, collects data for model training, and that’s a separate decision (OpenAI’s documentation on its crawlers).
- Perplexity surfaces and links websites through PerplexityBot, which its documentation says isn’t used to train foundation models (Perplexity’s documentation on its crawlers).
The practical takeaway: check your robots.txt. It’s easy to “block AI” on reflex, block the search crawlers along with it, and remove yourself from the answers.
The second change is about the kind of query. In August 2026 I checked which cleaning-industry searches triggered an AI Overview in Google. It appeared on 6 out of 6 question-style queries and on 0 out of 3 queries for a document template. That’s a small sample, but the direction is clear. AI answers the questions, while people still click through for ready-made documents and tools.
What still works from classic SEO
Almost all of it. In its guide to generative AI features, Google says this is still SEO and lists things you don’t need to do: create llms.txt files, chop content into small pieces for AI, rewrite text in a special style, or chase inauthentic mentions of your brand (Google’s guide to optimising for generative AI features).
| What | Why it still matters | What I check |
|---|---|---|
| Technical basics and indexing | a page outside the index can’t appear in results or in an AI Overview | indexing of key pages, server-rendered HTML, sitemap, canonical URLs |
| Core Web Vitals | Google recommends LCP within 2.5 s, INP under 200 ms and CLS under 0.1 | field data from real phones in Search Console |
| Helpful content | Google values first-hand experience over yet another list of tips an AI model could write | whether the page says something competitors don’t |
| Google Business Profile | local ranking depends on relevance, distance and prominence, and review count feeds prominence | categories, services, photos, reviews and replies |
| Brand signals | an AI can only recommend a company it can identify | a consistent name, address, tax ID and description on the site and in directories |
| Structured data | helps with rich results, but isn’t required to appear in AI features | whether the schema matches what’s visible on the page |
| llms.txt | a hint for AI assistants, not a ranking factor | whether the file states current facts |
The Core Web Vitals thresholds come from Google’s Core Web Vitals documentation, and the local factors from Google Business Profile help.
Here’s how that looks on my own site. In June 2026 Search Console reported a mobile INP of 205 ms for about 48 reefa.pl URLs, just over the threshold. I shipped four fixes the same day, and in July I cut the page’s JavaScript from 1,294 to 772 KB, a 40% reduction. I don’t have the post-fix field data on record yet, so I’m not presenting it as a win.
For brand signals, Reefa’s structured data carries a stable company identifier, tax ID, court register number and links to its profiles, and Google already has a Knowledge Graph entry for it. Since July 2026 reefa.pl has had an llms.txt file, but I have no measurement showing it changed anything. Google says such a file will neither help nor harm visibility in its search, because Google Search ignores it. I treat it as housekeeping for facts, not as a lever.
How a B2B company wins enquiries when customers ask AI
Pages that answer in the first paragraph
People and AI models read in a similar way: they look for the answer at the top. The first paragraph of a service page should say what you do, for whom, for how much and where, with no warm-up about mission and values. A paragraph like that can be quoted whole, and the customer knows straight away whether they’re in the right place.
The second rule is the same facts everywhere. On reefa.pl the FAQ structured data is generated from the questions visible on the page, so the schema says exactly what the text says. Every page draws on one approved set of company facts. That matters because AI repeats what it finds, old numbers included. On the blog, since 4 October 2026 the company facts are written into the instructions of the model that writes the articles.
Data nobody else has
An AI summary is stitched together from other people’s pages. If your page says what ten others say, a model has no reason to quote yours in particular. Your own data gives it one.
In August 2026 I put together an open dataset of cleaning time norms: 2,338 values from 168 sources in 13 countries. Eight AI agents collected it in parallel, one per market. A manual check of 173 rows found substantive errors in 1.7% of them, and I fixed every one. On 29 August I published the dataset on KtoSprząta.pl under a CC BY 4.0 licence: 460 categories in a CSV file, described with Dataset structured data.
A dataset like that produces findings nobody can copy from a competitor. For example, the time spent on ancillary tasks in cleaning isn’t a flat “+15%”. It’s 7 fixed minutes per shift plus a variable part: 48% at 100 m², 12% at 500 m² and 5% at 2,000 m². I also rebuilt the price calculators on reefa.pl on top of this dataset.
In a manufacturing or distribution business, that kind of data might be lead times, consumption norms, typical failures, or a cost comparison you already run every day. You don’t have to publish everything. One solid table nobody else has is enough.
Free tools instead of an e-book behind an email gate
A B2B buyer will use a calculator or a document generator more readily than hand over an email for a PDF. For reefa.pl and KtoSprząta.pl I built tools like these in two waves: 17 tools at the end of August 2026 and 18 more pages in early September. The tools menu on reefa.pl now lists 29 of them.
I follow two rules when building them:
- Every tool is a generator with an explanation next to it that answers the question directly. That comes from the measurement above: AI Overviews show up for questions, not for templates. The generator gives people a reason to click, and the explanation gives AI something to quote.
- The tool leads to contact naturally. The request-for-quote builder for housing associations lets you send the finished request to Reefa as well, and the building-inspection calendar export reminds you of each deadline 14 days ahead.
This site has two tools built the same way: the automation savings calculator and the AI readiness assessment.
Reviews and mentions beyond your own site
In July 2026 I ran a full visibility audit for Reefa. It showed the site wasn’t the bottleneck. What was missing happened elsewhere: Google Business Profile reviews, appearances in lists of companies, directory listings, and links that someone chose to give. AI builds its answer from many sources, so a company that only its own website talks about has a weaker chance of being recommended.
Google adds that chasing inauthentic mentions across the web isn’t as helpful as it might seem. Real ones count: customer reviews, listings in industry directories, articles where someone cites your data.
Measure enquiries, not traffic
Traffic is a supporting metric. A B2B company is paid by customers who sent an enquiry or picked up the phone, so that’s what I count first. Some AI recommendations leave no trace in analytics, because the customer types the company name into a search engine or simply calls.
What I set up in a company before changing any content:
- Every form records the page the enquiry came from and the traffic source.
- The form or the first call asks “How did you hear about us?”, and one of the options is “ChatGPT or another AI assistant”.
- Enquiries land in the CRM with their source, and once a month I check which pages brought them in.
- Search Console shows clicks and CTR. Google counts traffic from its AI features within overall search traffic in the Performance report, so I don’t treat it as a separate channel.
- Bing Webmaster Tools shows citations in Copilot. Reefa’s baseline from July 2026 is 161 citations over three months, all from one topic cluster (disinfection).
- A fixed list of customer questions, checked weekly across several assistants. For Reefa that’s 14 Polish questions in ChatGPT, Gemini and Google AI Overviews, with the first check scheduled for 22 October 2026.
I judge a change to a page title or description only after 35 days, and only if the page had at least 500 impressions and CTR moved by at least 0.4 percentage points. A shorter window mostly shows noise.
The reefa.pl numbers I have today:
| reefa.pl metric | Value | When | Type of number |
|---|---|---|---|
| Google clicks per day | from 8.6 to 11.2 (+30%) | 1–29 Jul 2026 vs 3 Jul 2026 | result after 30 days |
| Google CTR | from 0.88% to 1.20% | same period | result after 30 days |
| Keywords in the top 10, per Senuto | 36, the site’s highest ever | 29 Aug 2026 | measurement |
| Google Maps, “firma sprzątająca biura kraków”, 5 × 5 grid | average position 9.15, top 3 in 8 of 25 points | 30 Aug 2026 | measurement |
| CTR at an average position of 9.9 | 1.6% | 14 Oct 2026 | baseline |
| Tracked local keywords in the top 10, Kraków | 5 of 32 | 14 Oct 2026 | baseline |
| Copilot citations | 161 over three months | 4 Jul 2026 | baseline |
None of these numbers says how many more enquiries came in. I haven’t measured an increase in enquiries from this work yet, so I don’t claim one. That’s exactly why I set up enquiry tracking at the start, not six months in.
Honest limits
- Nobody controls AI answers. The same question gets different answers on different days and for different people, so every measurement is a sample and the trend over several weeks is what counts.
- Results take months. When I checked Reefa’s clicks after 30 days they had already risen, but reviews and mentions take longer to build.
- Content won’t replace reviews or reputation. If you compete locally with a company that has several times more reviews, your website alone won’t turn that around.
- Old facts come back. AI can repeat an outdated price or service range, so the same facts have to match on your site, your Business Profile and in directories.
- I don’t promise a spot in ChatGPT’s answer or a specific Google ranking. I promise work to a plan and measurement that shows whether it brings in enquiries.
Where to start in your company
- Check robots.txt: Googlebot, Bingbot, OAI-SearchBot and PerplexityBot should be able to reach your public pages.
- Rewrite the first paragraphs of your service pages so they say what you do, for whom, for how much and where.
- Complete your Google Business Profile and ask satisfied customers for reviews.
- Pick one data table or one tool that nobody in your industry has.
- Set up enquiry tracking by source before you change any content.
- Write down 10–15 questions your customers ask, and check monthly whom ChatGPT, Gemini and Google recommend.
This is how I work in my SEO and organic growth service (from PLN 7,000/month, excluding VAT). I describe how it went at Reefa in the reefa.pl organic growth case study.