Checked 14 October 2026
Eight things you can do today: five if you run a website, three if you ask AI engines for answers. Each one comes from a study on this site with a date, a sample and a CSV. None of them is a promise of traffic.
What this page will not claim
- That any of these changes traffic, citations or revenue. Nothing here was tested as a before-and-after.
- That an engine will behave the same way next month.
- Anything outside the samples and dates in the linked studies.
If you run a website
1. Block the training crawler, leave the fetcher open
What I measured: on 2 October 2026, 9 of the UK’s 10 biggest newsbrands blocked ClaudeBot, Anthropic’s training crawler. 5 of those 9 left Claude-User, the fetcher that reads a page when someone asks, unblocked: the BBC, Sky News, the Mirror, the Daily Express and Yahoo! (UK news). Across the sample, GPTBot was blocked by 7 and ChatGPT-User by 4. Full table and CSV.
The hack: if you want to stay out of training data but still be readable when a user asks about you, block the training tokens by name (GPTBot, ClaudeBot, CCBot, Google-Extended) and leave ChatGPT-User, Claude-User and Perplexity-User alone. That’s the split most of the UK’s biggest news sites have chosen. The difference between the three kinds of crawler is in Training bot, search bot, fetcher.
Check yours: open your own robots.txt and test each of the 15 crawlers in the method note. A whole-site Disallow: / in the group that names the crawler, or in the * group if nothing names it, counts as blocked.
Limits: robots.txt is a request, not a wall. It shows what a site asks for, not what crawlers do, and one day’s file isn’t a policy.
Full post: read it with the copy-paste block.
2. Put the date in the same sentence as the number
What I measured: on 10 October 2026 I asked four engines ten questions about “current” UK figures. 7 of 40 AI answers to “current” UK figure questions didn’t say when the figure was from. All 7 came from Google’s two engines. In one Google AI Mode answer the unemployment rate was right, but its period (May to July 2026) only appeared in a source snippet, not in the answer. Study and CSV.
A day earlier, on the purple pound: Both AI engines that answered gave £274bn, and neither said it was a 2020 figure. Study and CSV.
The hack: write the period into the sentence that carries the figure: “UK unemployment was 4.9% in May to July 2026”, not “4.9%” in a table with the date in a caption or a footnote. A figure and its date that sit in one sentence are harder to separate.
Limits: I didn’t test whether dating the figure changes what an engine repeats. The last sentence of the hack is my reasoning, not a result.
Full post: read it with the copy-paste block.
3. Write the customer into the page
What I measured: on 9 October 2026 I asked ChatGPT and Gemini for the best accounting software in the UK, with and without a persona. Three of the four finance-director answers recommended none of the products the no-persona answers named. The self-employed plumber got the same three brands but a different top pick. Study and CSV.
The hack: there’s no single list to rank in. Say on the page who the product is for, in the words that customer would type: “accounting software for a self-employed plumber”, not “SMB solution”. Then track your visibility one persona prompt at a time, dated.
Limits: two engines, three prompts, one evening, temporary chats. “Name three” forces a three-item list.
Full post: read it with the copy-paste block.
4. Keep facts in visible body text, not in markup
What I measured: on 12 October 2026 I ran one test page through four common open-source tools that turn HTML into text. Body paragraphs, headings and table cells were kept by all four. aria-labels, data- attributes, HTML comments, the meta description and JSON-LD were dropped by all four. Alt text was kept by two of the four. Text hidden with display:none was kept by three of the four. Study, test page and CSV.
The hack: if a fact matters (a price, a date, a name, a definition), put it in a sentence or a table on the page. Schema and meta descriptions do other jobs, but in this test they were the parts every tool lost. And don’t assume hidden text is hidden: most of these tools read the HTML without applying your styles.
Limits: four open-source tools on default settings, one static page. I don’t know what ChatGPT’s or Google’s own pipelines keep.
Full post: read it with the copy-paste block.
5. Naming a crawler lifts your catch-all block for it
What I measured: on 2 October 2026 the Sun’s robots.txt disallowed the whole site in its User-agent: * group, then gave OpenAI’s three crawlers groups of their own with path-level rules only. Result: GPTBot, OAI-SearchBot and ChatGPT-User were not blocked. The other 12 crawlers I checked were blocked by the catch-all. Notes on one file: the Sun.
The hack: a crawler follows the group that names it and ignores the * group (that’s the robots.txt standard, RFC 9309). So the moment you add User-agent: GPTBot to give it one rule, your site-wide Disallow: / stops applying to GPTBot. The Sun uses this on purpose to let OpenAI in. If you didn’t mean to, repeat Disallow: / inside the named group.
Check yours: for every crawler your file names, read only its own group. If Disallow: / isn’t in that group, the crawler isn’t blocked, whatever the * group says.
Limits: one file, one date. I can see how the Sun’s file is built, not why.
Full post: read it with the copy-paste block.
If you ask AI engines for answers
6. Ask when a “current” figure is from
What I measured: on 10 October 2026 I asked ChatGPT, Google AI Mode, Claude and Gemini ten UK figure questions with “now” or “today” in them. ChatGPT and Claude gave a period with all ten of their figures. Google AI Mode left it out four times and Gemini three times. Every headline figure matched the latest official release that day. The two wrong numbers I found were side figures: Gemini gave the youth unemployment rate as 14.4% (the ONS figure was 16.4%), and AI Mode quoted the 2025/26 National Living Wage. Study and CSV.
The hack: if an answer gives a “current” figure with no month, quarter or year, ask “when is that figure from?” and check the official source. Check the extra figures around the headline number too. In this test, that’s where both errors were.
Limits: one day, 40 answers, UK figures only. Perplexity wasn’t run because it asked me to sign up.
Full post: read it with the copy-paste block.
7. Say who you are in your first message
What I measured: in the 9 October accounting software test, one sentence about who was asking changed the shortlist. The finance director of a 250-person manufacturer got a different market from the generic question. ChatGPT ended all six of its answers with a question about my business, even when the prompt had already said. Study and CSV.
The hack: put your situation in the first message: your trade, your size, your budget, what you use now. A bare “what’s the best X?” gets you the answer for a customer who doesn’t exist.
Limits: I tested whether a persona changes the answer, not whether it makes it better.
Full post: read it with the copy-paste block.
8. Ask twice before you trust a shortlist
What I measured: in the same test, every prompt ran twice in fresh chats. Sage Accounting was named in one generic ChatGPT run and not in the other. In both plumber runs, ChatGPT’s top pick was FreeAgent. Study and CSV.
The hack: before acting on an AI shortlist, ask again in a new chat with the same wording. What appears both times is the steadier part of the answer. What swaps out between runs wasn’t settled.
Limits: two runs per prompt is enough to see answers move, not to measure how often they do.
Full post: read it with the copy-paste block.
A new hack goes on this page only when a dated study on this site backs it. If a later snapshot changes a number, the hack gets a dated update line. It isn’t quietly rewritten.