AI visibility, and the email you can lawfully send
Two bodies of evidence sit behind this article, and both were read from the primary pages. The first is how AI assistants describe and cite small businesses. The second is what the statutes and regulators in Ireland and the UK say about direct business email. Two findings carry the piece. No operator promises a citation, and the controls are narrower than the market claims. Cold business email is lawful in both countries, and the two arrive by different routes. What has been prosecuted in Ireland is an ignored opt-out, not a first approach. The piece is written for the owner of a small Irish or UK firm who would rather use both channels than guess at either.
What we measured, and what we did not
Two bodies of evidence sit behind this article, and the first is an open register of factual claims. Fact Provenance republishes as W3C PROV-O records any sentence on a published page that carries a number, a date, an attribution or a superlative. Its statistics endpoint counts 14,870 claims drawn from 2,692 pages across 23 source domains, and 1,831 of those pages carry a pointer to a source the page itself published. A record lists the claim, the page it came from, that page's publication date, a SHA-256 digest of the page as fetched, the fetch time, the extraction method, the agent that ran it, and a verification status. None of the records is evidence that a fact is true, and the register says so plainly: "Unverified, every record here is indexed from a published page. 'Cited source' is the pointer the page itself published; the quote has not yet been matched against it. Nothing in this index claims a fact is true."
Documentation is the second body: the operators' crawler and search pages, the deliverability rules Google and Microsoft publish and enforce, the technical standards, and the legislation and regulator guidance on marketing email. Every quotation was read on the live page on 26 September 2026, and the URLs and page dates sit in the sources list. Three limits apply. We ran no controlled citation test across assistants, so the test we describe is one a reader runs on their own business. We sent no campaign, so the thresholds are the published ones and not our own inbox placement. The register reads page text only, so prices rendered by script are missed.
What an AI assistant actually sees of your business
A machine sees what it can fetch over HTTP, in text, at the moment it goes looking. Your premises, your van, your sign, and the four paragraphs of your services page that appear only after a plugin runs are all invisible to it. That sounds obvious. Then you count how much of a small business lives only in the physical world, and how confidently the gap gets filled.
The advice oversells how much gets read. Google documents that a page must be "indexed and eligible to be shown in Google Search with a snippet" to qualify as a supporting link in AI Overviews or AI Mode, and that "indexing and serving isn't guaranteed" even for a page that meets every requirement. Third parties are read on the same footing as your own site, and that is where small firms lose control of their facts. A directory listing, a map pin, a review page and a register entry are all fetchable statements about you, and if one of them is the only public record of your opening hours, that one becomes the record. The page you want is the one that states your facts plainly enough to be copied, not the one that out-volumes the reference sites.
One exception bends the other way. Some of what an assistant reads arrives because a person asked it to look. OpenAI states that ChatGPT-User "is not used for crawling the web in an automatic fashion", and that "because these actions are initiated by a user, robots.txt rules may not apply". A reader who pastes your URL into a chat window is not the visitor a crawler is.
Cited, summarised wrong, and absent are three different problems
Owners read "I am not in the AI answer" as one problem with one fix. It is three, and each one takes different work.
Absent. A machine can only fetch what a page states plainly: what you do, where, for whom and at what price. A page that never says it cannot be quoted. This case is the most common and the most fixable: write plain sentences on your own pages, and keep the third-party pages about you current.
Summarised wrong. Facts about you exist but are stale, or they belong to someone else: an address from a directory nobody updated, a service you dropped in 2023, a competitor's price in a shared table. The assistant is repeating a published sentence that was once true. It is not inventing. The register is built around this problem, because every record carries the URL it came from and the date that page was fetched, so a reader can see whether a sentence came from an old page or from today.
Cited. This splits in two. A citation is a visible link to your page inside an answer. A mention with no link is not a citation and cannot be audited, because there is no page to check. Microsoft is the only operator we found publishing a report for the distinction: its AI Performance report shows "how often a specific URL from your site is visibly cited in an AI-generated answer", with the caveat that it "does not measure rankings, authority, performance, or importance. It simply shows which content was cited."
A citation is also less stable than a ranking. A search result is a position on a page produced by a system with a published interface. An answer is generated per question, so two people asking slightly different questions in the same week can get different links, or none. The register's record shape is worth copying for your own published figures: the claim, the page it is on, the date of that page, and what was fetched.
How retrieval and citations work, as far as the operators document it
How an answer is assembled is documented by nobody outside the operators. What is published is eligibility and control, and those two are worth keeping apart.
Google states the eligibility test in one sentence: "To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements." It then discards the industry's invented requirements in two more: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add." One control does exist, on the same page: because "AI is built into Search and integral to how Search functions", robots.txt rules for Googlebot are the control a site owner has.
OpenAI documents two crawlers for two purposes, and the split is the point. GPTBot crawls content that may be used in training its models. OAI-SearchBot "is used to surface websites in search results in ChatGPT's search features", and a site opted out of it "will not be shown in ChatGPT search answers, though can still appear as navigational links". Allowing one and disallowing the other is a supported choice. OpenAI adds that a robots.txt change can take roughly 24 hours to apply, and publishes IP ranges per crawler for an owner who would rather check a visitor than trust a string. Anthropic documents three tokens for three jobs, ClaudeBot, Claude-User and Claude-SearchBot, and says its bots "respect 'do not crawl' signals by honoring industry standard directives in robots.txt".
The protocol shields less than it is often described as doing. RFC 9309 says crawlers "are requested to honor" robots.txt rules and then states plainly: "These rules are not a form of access authorization." A rule is a request that well-behaved crawlers follow, and a path listed in robots.txt is public, so the file cannot hide a page.
Crawler controls, and what each one actually does
A robots.txt file holds a set of groups, each naming a product token and listing paths the named crawler should not request. RFC 9309 sets out the matching: the token matches as a substring of the crawler's identification string, the most specific rule wins, and a URI "MUST be considered allowed" when no rule matches. Deleting the file is therefore not a block. An empty one is not a shield. The table below lists the controls the operators publish.
| Control | Published by | What allowing it permits | What disallowing it stops |
|---|---|---|---|
| Googlebot | Crawling for Google Search, including the AI features built on that index, plus Images, Video, News and Discover. | Indexing for Search, and with it eligibility as a supporting link in AI Overviews or AI Mode. | |
| Googlebot-Image, Googlebot-Video | Images, Video and Discover, and the logos and favicons shown in Search features. | Those surfaces, even where the main Googlebot token is allowed. | |
| Google-Extended | Use of crawled content for training future Gemini models, and for grounding in Gemini Apps and on Vertex AI. | Training and grounding only. It "does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search". | |
| GPTBot | OpenAI | Crawling that may feed model training. | Use of the content in training, without touching the search crawler. |
| OAI-SearchBot | OpenAI | Surfacing the site in ChatGPT's search features. | The site will "not be shown in ChatGPT search answers, though can still appear as navigational links". |
| ChatGPT-User | OpenAI | Fetches triggered by a person asking ChatGPT to look at a URL. | Little in practice: the actions are user-initiated and robots.txt "rules may not apply". |
| OAI-AdsBot | OpenAI | Checking pages behind ads submitted to ChatGPT. | Ad validation for those URLs. |
| ClaudeBot, Claude-User, Claude-SearchBot | Anthropic | Crawling, user-initiated fetching and search, as three separate jobs. | The named job, and nothing outside it. |
| bingbot, AdIdxBot | Microsoft | Bing's index and the Copilot and grounding surfaces that draw on it, and advertising-related crawling. | Bing indexing, and with it the citation surfaces Microsoft reports on. |
The identification strings matter more than the tokens, since a token is only a substring match. Google publishes the following, with a placeholder for the browser version:
Mozilla/5.0 (Linux; Android 6.0.1; Nexus 5X Build/MMB29P) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/W.X.Y.Z Mobile Safari/537.36 (compatible; Googlebot/2.1; +http://www.google.com/bot.html)
Mozilla/5.0 AppleWebKit/537.36 (KHTML, like Gecko; compatible; Googlebot/2.1; +http://www.google.com/bot.html) Chrome/W.X.Y.Z Safari/537.36
Google also publishes a placeholder for the browser version, noting that "Chrome/W.X.Y.Z in the user agent strings in the list is a placeholder that represents the version of the Chrome browser used by that user agent", and tells anyone filtering logs to "use wildcards for the version number rather than specifying an exact version number". The stable fragments carry the URL in brackets: Googlebot/2.1, and Googlebot-Image/1.0 or Googlebot-Video/1.0 for the media crawlers. OpenAI publishes four strings: OAI-SearchBot/1.4; +https://openai.com/searchbot, GPTBot/1.4; +https://openai.com/gptbot, ChatGPT-User/1.0; +https://openai.com/bot and OAI-AdsBot/1.0; +https://openai.com/adsbot. Microsoft publishes two strings: Mozilla/5.0 (compatible; bingbot/2.0; +http://www.bing.com/bingbot.htm) and Mozilla/5.0 (compatible; AdIdxBot/2.0; +http://www.bing.com/bingbot.htm). No string proves identity. Google states outright that "The HTTP user agent string can be spoofed", which is why the operators also publish IP ranges.
llms.txt and ai.txt are not standards, and no operator says it reads them
Two file names circulate in AI visibility advice, and neither is a standard. No operator we checked documents using either. llms.txt is a proposal by Jeremy Howard, published on 3 September 2024 and revised on 10 August 2026, and describes itself as "A proposal to standardise on using an /llms.txt file". Google addresses machine-readable AI files directly: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them." We found no crawler documentation from OpenAI, Anthropic or Microsoft stating that it reads a third-party llms.txt. ai.txt has no authoritative specification at all: the only vendor material we found is a generator page run by Spawning, which says of its own file that "many AI miners respect the rules set in ai.txt files, some may not adhere to them".
Structured data and consistent public facts
Structured data is worth doing for the reason Google gives, not the reason the market gives. Google describes it as "a way to provide explicit clues about the meaning of a page to search engines", with two limits: it must be "representative of the main content of the page" and describe "the visible content of a page", and you should not "create empty pages just to hold structured data". Markup cannot assert a fact the page does not show, and it plays no documented part in AI feature eligibility. Add it for the search features that consume it, and treat it as a bonus, not a control.
Consistency is the one lever that moves. A machine reading about you assembles facts from whichever pages it fetched, so your address survives in the version on the page that best answers the question asked. Publish one canonical facts page in the same wording as everywhere else: what you do, where you are, who you work for, your price range or hourly rate, your hours, your service area, and the date you last checked it. Then bring the surfaces you control into line with it, the site footer, the email footer, the Google Business Profile, and the two or three directories that rank for your trade. For figures, copy the record shape the open register uses, described in our own piece on the register of provenance: one number, the page it appears on, that page's date, the source behind it, and the date of your last check.
A ten point audit you can run this week
Everything in this section is free, and most of it takes one afternoon. Record what you find, because the point is to compare it against next month.
- Write five prompts down before you run them: what the business does, what it charges, whether it is open on a Saturday, who does your trade in your county, and its name. Save the exact wording, the model and the date.
- Run each prompt and note three columns: cited, mentioned, absent. A citation is a visible link to your own page inside the answer. Where no click-through to one of your pages exists, it is a mention, and a mention cannot be checked, because there is no source to inspect.
- Fetch your own site the way a crawler does, with
curl -sL https://yourdomain.ie/ -o page.html, then search that file for a sentence you know is on the page. If the sentence is absent, script renders it after load, and the page may read as empty to anything that fetches instead of renders. - Read your robots.txt file with
curl -s https://yourdomain.ie/robots.txtand check it against the table above, so the tokens you meant to block are blocked and the ones you need are allowed. A missing file allows everything. A strayDisallow: /hides the whole site. - Confirm that the pages you want quoted are indexed and snippet eligible, starting with a
site:yourdomain.iesearch. A page that cannot be indexed cannot be cited, and a page markednoindex,nosnippetordata-nosnippetwill not be quoted even once it is found. - Audit the pages other people wrote about you: your Google Business Profile, the directories in your trade, review sites and your company register entry. Note the URL, what it says, whether it is still true, and the date you looked.
- Give one published number a provenance record: the claim, the page, that page's date, the source behind it, and the date of your last check.
- Publish one canonical facts page in plain text with a "last updated" line, and link it from your home page. Three pages that disagree are worse than one page slightly out of date.
- Prepare the outreach half the same afternoon: for every address, record the URL you took it from and the date, plus whether the recipient is a limited company or a sole trader. In the UK, that column decides between a lawful email and a breach.
- Set up the sending domain and monitoring before the first message goes out: the subdomain, SPF, DKIM, DMARC, Google Postmaster Tools, Microsoft SNDS and JMRP, and one suppression list screened against every campaign.
If you would rather have the facts page, the prompt record or the list built for you, that is work we do, and the checklists in our shop cover the do-it-yourself version.
When direct email is the right move
Direct email suits a market whose buyers you can name: ten garden centres within an hour's drive, forty accountancy practices in two counties, the eleven contractors on a tender list. It does not suit consumer lists, bought lists, or any plan that leans on volume, because volume is what turns a legitimate message into a complaint rate, and a complaint rate closes your channel.
The test is one sentence: "I am writing to you because". If only something generic finishes it, the email is not ready. If something specific and true about that business finishes it, published by that business, the message has a reason to exist and the rest is craft.
What the law allows in Ireland and the UK
Both countries permit unsolicited marketing email to businesses, but they say so differently, and the wording matters when you decide what to record about a list.
Ireland: a consent rule with a commercial activity carve-out
Regulation 13 of the ePrivacy Regulations, S.I. No. 336 of 2011, is the operative provision. Regulation 13(1) prohibits sending "to a subscriber or user who is a natural person an unsolicited communication for the purpose of direct marketing" by electronic mail "unless the person has been notified by that subscriber or user that he or she consents to the receipt of such a communication". Read on its own, that reaches business email, because an employee is a natural person.
Regulation 13(2) settles the question, and is worth quoting in full: "Notwithstanding paragraph (1) and subject to paragraph (4), the use of electronic mail to send an unsolicited communication for the purpose of direct marketing to a natural person does not include an electronic mail to an email address that reasonably appears to the sender to be an email address used mainly by the subscriber or user in the context of their commercial or official activity and the unsolicited communication relates solely to that commercial or official activity." Two conditions sit inside it: the address must reasonably appear to you to be used mainly for commercial or official activity, and the message must relate solely to that activity.
Regulation 13(4) covers everyone else, prohibiting unsolicited marketing email "to a subscriber or user other than a natural person, where the subscriber or user has notified the person that the subscriber or user does not consent to the receipt of such a communication". Regulation 13(12) forbids email that "disguises or conceals the identity of the sender" or that "does not have a valid address to which the recipient may send a request that such communication shall cease". Regulation 13(13) makes a breach an offence and adds that "the sending of each unsolicited communication or electronic mail ... constitutes a separate offence", which is why the risk scales with list size and not with campaign count. Regulation 13(11) allows your own similar products to be marketed to a customer for 12 months after a sale if they could easily object at the time and in every message. The Irish dividing line is natural person or not.
The Data Protection Commission is clear on ignored opt-outs. Its FAQ says: "It is a criminal offence to send electronic direct marketing communications to a person who has not consented to receiving such communications either by opt in consent or, in the case of an existing business relationship with the person, by means of an opt out", and that "a valid address to opt-out must be included with each marketing communication". What it does not say, on any live page we could reach, is that company email addresses fall outside the consent rule: its current FAQ pages contain no mention of businesses, corporate subscribers or B2B email. The statement we can quote comes from an Internet Archive snapshot dated 20 June 2019: "In the case of businesses, messages can be sent until such time as the sender is asked to stop and any subsequent messages from that sender would then be an offence." Treat that as historical, not live.
The enforcement record says more than the live guidance does. Of the 19 entries in the DPC's electronic direct marketing case studies index, one concerns a business email address: the prosecution of Viking Direct (Ireland) Limited, where a business owner had asked to be removed from the marketing list and further emails continued. At Dublin Metropolitan District Court on 14 May 2018 the company "entered a guilty plea to one charge of sending an unsolicited marketing email to a business email address in contravention of Regulation 13(4)", and at sentencing on 11 June 2018 the court applied section 1(1) of the Probation of Offenders Act in lieu of a conviction and fine, and the company covered the DPC's costs. The case was about a refusal that was ignored, not a first approach.
Two general rules sit on top. Sending to a named individual's business address is processing personal data, so a lawful basis is needed, in practice Article 6(1)(f) of the GDPR, covering processing "necessary for the purposes of the legitimate interests pursued by the controller or by a third party" except where those interests are overridden by the individual's rights and freedoms. Article 21(3) is absolute: "Where the data subject objects to processing for direct marketing purposes, the personal data shall no longer be processed for such purposes." In Ireland, section 58 of the Data Protection Act 2018 includes direct mailing in that right.
The United Kingdom: the rule simply does not reach companies
PECR goes the other way. Regulation 22(1) applies the marketing rule "to the transmission of unsolicited communications by means of electronic mail to individual subscribers", and regulation 22(2) requires prior consent from those recipients unless the soft opt-in in regulation 22(3) applies, which needs details obtained "in the course of the sale or negotiations for the sale of a product or service", marketing of "that person's similar products and services only", and a simple means of refusing at collection and in every message. Regulation 23 forbids marketing email that conceals the sender's identity or that gives no "valid address to which the recipient of the communication may send a request that such communications cease".
The ICO states that scope limit plainly: "The PECR rule on direct marketing by electronic mail does not apply to corporate subscribers. For example, this means you can send B2B direct marketing emails or texts to any corporate body. You do not need their consent under PECR to send such messages." Two duties remain, not concealing your identity and giving a valid address for opting out, and the ICO settles the opt-out point in the same passage: "PECR does not say that you must comply with a corporate subscriber's opt-out in the context of electronic mail. However ... it serves little purpose to continue to send such marketing to a corporate subscriber who has asked you not to. Therefore, you should comply with a corporate subscriber's opt-out request."
Regulation 2 defines a corporate subscriber to cover companies, corporations sole, limited liability partnerships, Scottish partnerships, some government bodies and other bodies corporate distinct from their members, and the ICO notes that "the email address or telephone number of an employee at a corporate body would constitute a corporate subscriber ... because the 'subscriber' is their employer". Sole traders and other partnerships fall on the other side: they are "classed as 'individual subscribers' and PECR treats them the same as individuals", so you may only email them "if they have specifically consented, or the 'soft opt-in' applies". PECR does not define individual subscriber, so that classification rests on the ICO's interpretation, and where the answer is unsure, the advice is to treat the details as belonging to an individual subscriber.
A named business contact still sits under the GDPR overlay. Where PECR does not require consent, the ICO calls it "likely that legitimate interests will be the appropriate lawful basis", subject to identifying the interest, showing necessity and balancing against the individual's rights, and that a contact found through a public website still needs a lawful basis and "providing the individual with privacy information". If the contact objects, the right is absolute: "This right is absolute and there are no grounds for you to refuse." Two dates matter: the guidance page carries a notice that it is being reviewed because of the Data (Use and Access) Act 2025, and the parent PECR guidance records a latest update of 20 August 2025, with regulation 22 amended with effect from 5 February 2026.
Where the three positions differ, and what is unsettled
The EU floor explains why they diverge. Article 13(1) of Directive 2002/58/EC requires prior consent for marketing email, but Article 13(5) confines that to natural persons and asks member states only to ensure that "the legitimate interests of subscribers other than natural persons with regard to unsolicited communications are sufficiently protected". Business email was left to national law from the outset, and the two countries used that freedom differently: Ireland as a carve-out with conditions, the UK as a scope limit that never engages. The proposed ePrivacy Regulation that would have replaced the Directive is logged on EUR-Lex as "Proposal withdrawn", by Commission withdrawal dated 06/10/2025, so the Directive remains the instrument.
Four things are genuinely unsettled or unverifiable, and they belong in your notes and not your marketing copy. The Irish test turns on whether an address reasonably appears to be used mainly for commercial activity, and no published Irish decision settles where that phrase ends. The citable Irish authority for cold business email being lawful is thin, and the page that said it is no longer publicly readable. In the UK, the classification of sole traders as individual subscribers rests on the ICO's interpretation, and the ICO's electronic mail guidance is under review following the 2025 Act. The ICO's own warning about cross-border sending applies to an Irish firm writing into the EU: "some of them are stricter than the UK regulations, especially for marketing to companies".
The United States runs a different regime, and it is worth knowing if you email American businesses. The FTC's guide to the CAN-SPAM Act states that it "makes no exception for business-to-business email", requires accurate headers and subject lines, identification of the message as an advertisement, a valid physical postal address and a working opt-out, and lists penalties of "up to $53,088" for each separate email in violation.
What separates a legitimate email from spam
Spam is not so much a legal category as a judgement made in a second by a recipient, and in a moment by a filter. Four things decide it, and the law cares about the same four.
Relevance. The first line should give the reason for this business in terms only that business could recognise: a page they published, an event they announced, a gap visible on their own site. If the line would work unchanged for any other recipient, it is a mail merge.
Source of the address. Record the URL and the date for every address, because that record is what you check a complaint against, and it makes the Irish commercial activity test and the UK corporate subscriber column answerable instead of arguable. The ICO frames both halves honestly: a person whose "business contact details are publicly available on their employer's website may well expect to receive contact from other businesses", while still requiring "a lawful basis for the processing and providing the individual with privacy information".
Opt-out. Every message carries a valid address for refusal, refusals are actioned immediately, and every future list is screened against them. The ICO tells senders to "keep a 'do not email or text' list of any businesses that object or opt out, and screen any new marketing lists against that", and Google treats a sender whose unsubscribes are not honoured within 48 hours as failing its requirements. An opt-out that takes a week, or that needs a login, is a defect.
Honest sender identity. A real firm, a real reply address, a real postal address, and a subject line that describes what is inside. Regulation 13(12) in Ireland and regulation 23 in the UK make concealment an offence in its own right, independent of consent.
Underneath those four sits the lawful basis, and you should be able to write it in two sentences: the interest you are pursuing, why contacting this person is necessary for it, and why their interests do not override yours. If you cannot write those sentences, the campaign is not ready, and the sentence you cannot write is usually the one about necessity.
List hygiene and deliverability setup
Three records and one header set carry the technical half. SPF, in RFC 7208, lets a domain owner "explicitly authorize the hosts that are allowed to use their domain names, and a receiving host can check such authorization". DKIM, in RFC 6376, "permits a person, role, or organization that owns the signing domain to claim some responsibility for a message", validated "through a cryptographic signature and by querying the Signer's domain directly to retrieve the appropriate public key". DMARC, in RFC 7489, "authenticates use of the RFC5322.From domain by requiring that it match (be aligned with) an Authenticated Identifier", and its policy tag chooses the receiver's action against a mismatch: p=none "requests no specific action be taken", quarantine treats the message "as suspicious", and reject asks for rejection during the SMTP transaction.
Google publishes its thresholds, which makes them the easiest to work to. Since 1 February 2024 all senders to Gmail accounts must authenticate with SPF or DKIM, publish valid forward and reverse DNS records, use TLS, keep the spam rate below 0.3% and format messages to RFC 5322. Above 5,000 messages a day, SPF and DKIM must both pass, a DMARC record must exist with a policy that "can be set to none", the From header must be aligned, and marketing messages must support one-click unsubscribe with a visible link in the body. The numbers that decide your fate sit in the sender guidelines FAQ: keep the user-reported spam rate "below 0.1%" and "prevent spam rates from ever reaching 0.3% or higher", because since June 2024 "bulk senders with a user-reported spam rate greater than 0.3% will be ineligible for mitigation". Its consequences table lists four failures that each remove delivery support: a spam rate above 0.3%, a missing DMARC record at a minimum of p=none, marketing messages without one-click unsubscribe, and unsubscribe requests not honoured "within 48 hours". From November 2025 Google is "ramping up its enforcement on non-compliant traffic".
Microsoft enforces a parallel requirement from 5 May 2025: a domain sending more than 5,000 emails a day to Outlook.com "must fully be compliant with SPF, DKIM, and DMARC". A non-compliant message goes "to the junk folder" first and is then rejected with the error "550; 5.7.515 Access denied, sending domain [SendingDomain] does not meet the required authentication level". Microsoft recommends keeping bulk mail on a custom subdomain and exposes complaint and trap data through SNDS and the Junk Email Reporting Program. Unlike Google, it publishes no numeric complaint or bounce threshold that we could find.
Warm-up advice needs a label. Google publishes no warm-up schedule at all. Microsoft documents one, in the documentation for its Dynamics 365 marketing product, not as an Outlook.com sender requirement: "Achieving maximum deliverability takes four to eight weeks, depending on the targeted volume and engagement", with the first two weeks sent to subscribers active in the past 30 days. Treat that as a pattern from a product manual, not a receiver requirement. The hygiene rules carry no such caveat: remove hard bounces at once, without retrying them, verify addresses before the first send, re-verify anything older than a few months, keep one suppression list screened against every campaign, and hold the volume steady instead of spiking it.
What to write: a worked example
None of the businesses below exist. Kilbeggan Sign & Print is a fictional six-person sign shop in County Westmeath, Barrow Valley Garden Centre a fictional family garden centre in County Carlow, and Wrenfield Lane Nursery a fictional grower in Norfolk.
Two places supplied Kilbeggan's list, recorded per row. Barrow Valley's address came from the contact page on the centre's own website on 22 September 2026, and the row notes that it is a limited company. Wrenfield Lane's came from a trade association directory where the nursery lists itself, and the row flags it as a partnership, which in the UK means treating it as an individual subscriber. That column decides whether the email is lawful.
What Kilbeggan sent is below. It is one message written for one recipient, not a template to run at volume.
Subject: Sign for the Carlow road entrance at Barrow Valley
Hello Aisling,
I run Kilbeggan Sign & Print, a six-person sign shop in Westmeath. I am writing because your website shows a new display area opening this spring, and the entrance sign on the Carlow road is still the old panel in the photograph on your news page.
We cut, weld and fit signage for garden centres and farm suppliers, and we work within a day's drive of Carlow. Our own page lists what jobs cost and how long they take, including the entrance panel most centres your size choose.
If it would help, I will measure the entrance and send one price for a replacement panel and one for a full entrance set, with no obligation either way. If the timing is wrong, reply and I will not write again.
You are receiving this because your business address is published on your own website. If you would prefer not to hear from us, reply with "stop" and we will remove you the same day.
Fergus Nolan
Kilbeggan Sign & Print, Unit 4, [street address], Kilbeggan, County Westmeath
[phone number] | [reply address]
Read it against the four requirements. The relevance is specific and checkable. The address has a recorded source with a date. The opt-out works by reply, is actioned the same day, and says so. The sender is a real firm with an address, a phone number and a reply address that reaches a person who answers, and nothing is attached on a first contact.
One follow-up a week later, adding something useful in place of repeating the request, and then stop. A refusal is a suppression-list entry, and in Ireland an ignored refusal is the only thing in this area that has been prosecuted.
Kilbeggan ran the AI half of the audit the same week, and the two interlock. It published one dated facts page carrying what it makes, the counties it covers and its typical price range, then ran the five prompts and found the assistants knew it existed but had its 2019 workshop address, taken from a supplier directory nobody had updated. It corrected the directory, re-dated the facts page, and re-ran the same prompts a fortnight later, keeping both results in the same spreadsheet.
What to measure, and how to tell it is working
On the visibility side, count prompts and not impressions, in three columns: cited with a clickable link, mentioned with no source, and absent. Repeat monthly and keep the dates. A Bing Webmaster Tools account adds a provider-side number, "Pages Cited in AI Answers", alongside its documented caveat that it "does not measure rankings, authority, performance, or importance".
On the sending side, four numbers and one judgement: the spam rate in Google Postmaster Tools, which belongs under 0.1% and must never approach 0.3%; the complaint and trap data in Microsoft SNDS and JMRP; the hard-bounce rate, which should fall as the list is cleaned; and the reply rate broken down by the source of each address, which is how you learn that a directory list beats a scraped one. The judgement is whether the replies come from businesses that look like your existing customers, because replies from outside your profile mean the list is wrong, and no copywriting fixes a wrong list. Working, after a month, looks modest and checkable: a handful of conversations, a falling bounce rate, no complaints, and a dated record showing the assistants now describe your address and price range as you publish them.
What this means for you
Seven actions, cheapest first. Run the five prompts today and write down what you get, because the record is the baseline and nothing reconstructs it later. Publish one dated facts page in plain text and link it from your home page and your email footer. Read your robots.txt and decide deliberately what you are blocking. Fix the three third-party pages that describe you most often, starting with the one being quoted. Set up the subdomain, SPF, DKIM, DMARC and Postmaster Tools before the first message leaves. Start at 20 to 50 researched emails a week, from addresses whose source you recorded per row, with one follow-up and then silence. Keep two records, the prompt results and the suppression list: one shows what the assistants say, the other keeps you lawful.
None of that needs a budget or a tool subscription, and the pieces behind it sit in our insights section, with the services around it described in what we do. Start in the contact form with the trade you sell to and the counties you cover, and we will tell you which of the seven steps matters most for your market. If you spend money on any of it, spend it on list research and copy, because those decide whether the channel works.
What we do not know
No citation can be guaranteed and neither can inbox placement. Anyone promising a position in ChatGPT, or a guaranteed arrival in the primary inbox, is selling something outside their control, and that includes us. The work is what can be promised: published facts, a recorded source per contact, working authentication, an honoured opt-out, and a dated record of what was measured.
Beyond that, six things are open. The operators do not publish how an answer is assembled or which source wins for a question, so eligibility and control are all that can be described. Anthropic publishes tokens but no user-agent strings, and we could not confirm Perplexity's citation policy from its own help pages. Microsoft publishes no numeric complaint or bounce threshold, unlike Google's 0.1% and 0.3%, and Google publishes no warm-up schedule, so the four-to-eight-week figure comes from Microsoft's marketing-product documentation. The Irish regulator's live pages contain no statement of the business email position, and the wording that states it is a 2019 snapshot of a page that is no longer publicly readable. Regulation 13(2) turns on whether an address reasonably appears to be used mainly for commercial activity, with no published Irish decision settling its edges.
Two closing cautions. Operator pages change: the Google crawler reference was last updated on 14 July 2026 and the AI features page on 10 December 2025, so re-check the source, and do not trust our dates. The claim index used in the first half is a prototype whose records are explicitly unverified: it shows where published claims come from, not that any of them is true.
Sources
Every quotation was read from the live page on 26 September 2026, unless the entry says otherwise.
- S.I. No. 336 of 2011, regulation 13: consent, the commercial-activity carve-out, opt-out, the soft opt-in, identity, and the offence.
- S.I. No. 333 of 2011, regulation 2: the definition of subscriber.
- Data Protection Act 2018, section 58: direct mailing.
- Data Protection Commission FAQ: the offence, and the opt-out address.
- DPC case study, Viking Direct: the only business email prosecution.
- EUR-Lex procedure 2017_3: the withdrawal of the ePrivacy Regulation.
- Directive 2002/58/EC, Article 13: the consent rule, and the natural-persons limit.
- Regulation (EU) 2016/679, Articles 6 and 21: legitimate interests, and the right to object.
- PECR regulation 22, regulation 23, regulation 2.
- ICO, business-to-business marketing: corporate subscribers, sole traders, and the opt-out duty.
- ICO Guide to PECR, electronic mail marketing: the do-not-email list.
- ICO Guide to PECR, electronic and telephone marketing: updated 20 August 2025.
- FTC, CAN-SPAM compliance guide: the opt-out regime, and the penalty per email.
- Google email sender guidelines, sender guidelines FAQ, Postmaster Tools.
- Microsoft: Postmaster policies, high-volume sender requirements, the 2 April 2025 announcement, SNDS, warm-up.
- Standards: RFC 7208, RFC 6376, RFC 7489, RFC 8058, RFC 2369, RFC 9309.
- Google: common crawlers (14 July 2026), crawler overview (12 June 2026), AI features (10 December 2025), AI optimization guide, structured data, robots.txt.
- OpenAI crawler overview: four bots, and their IP ranges.
- Anthropic crawler article: three tokens.
- Bing: crawlers, webmaster guidelines, AI Performance report.
- llms.txt (revised 10 August 2026) and the ai.txt generator.
- Fact Provenance claim index: the statistics endpoint and record 6445.
Revision log
26 September 2026: first draft, merging the two earlier short pieces on AI visibility and business email into one article, with every source re-read from the live page the same day.