Because it is the least rewarding writing task in marketing. A few hundred pages, a few hundred nearly identical summaries, and no visible payoff for any single one. It is the definition of work that should be automated, which is exactly why AI got pointed at it first.
The avoidance shows in the data. The 2025 Web Almanac from HTTP Archive found that "Title tags are present on 98.6% of desktop pages and 98.5% of mobile pages", while "Meta descriptions are present on 67.7% of desktop pages and 67.2% of mobile pages". Roughly a third of the web simply skips them.
So the question is not really whether to use AI here. It is how to use it in a way that produces something better than the blank field it replaces. Our answer is yes, with a process, and the process matters more than the model.
It offers Google a summary to display under your title in search results. It is not a ranking factor, and it never has been. What it influences is whether someone clicks, which is a different thing that still shows up in your traffic.
That distinction sets the right expectation. Rewriting every description on a site will not move your positions. It can move your click-through rate, and on a page already ranking in the top handful of results, click-through rate is one of the few levers left that does not require earning links or rewriting the content.
Treat the description as ad copy for a listing you cannot otherwise control. The reader is comparing ten results in about two seconds. Your job is to make the choice obvious for the people you actually want, which is a copywriting problem rather than an SEO one.
Often not, and this is where most of the frustration comes from. Google's documentation is explicit: "Snippets are primarily created from the page content itself. However, Google sometimes uses the meta description HTML element if it might give users a more accurate description of the page than content taken directly from the page."
Read that carefully, because the logic is the opposite of how people usually describe it. Google is not overriding your description out of spite. It is choosing whichever text better matches the specific query someone typed, and a generic description loses that comparison to a well-matched sentence pulled from your page.
The practical takeaway is that a great description gets used more often. A vague one written to cover every possible visitor gets replaced almost every time, because there is always a paragraph on the page that fits a specific query better. Writing more specifically is how you win the slot, which is also the argument we made in our guide to title tags and meta descriptions.
No, and Google's own documentation goes further than most people realise. On sites with large numbers of pages, its snippet guidance states that "programmatic generation of the descriptions can be appropriate and are encouraged", with the condition that they remain "human-readable and diverse".
That is a genuinely permissive position, and it has been Google's stance since long before current AI tools existed. Programmatic generation from a database was always acceptable. Generating from a language model is the same idea with better output, provided the two conditions still hold.
The limit sits elsewhere, in Google's helpful content guidance, which warns that using automation "to produce content for the primary purpose of manipulating search rankings" violates its spam policies. Generating an accurate summary of a page you wrote is not that. Generating a thousand pages of filler is. The line is about intent and accuracy, not about which tool typed the words.
Shorter than most sites write them. Google states plainly that "There's no limit on how long a meta description can be, but the snippet is truncated in Google Search results as needed, typically to fit the device width." So there is no hard cap, only a display cut-off that varies by device.
The Web Almanac data shows how far practice has drifted from that. The median meta description is "40 words" and "274 characters on both desktop and mobile", and the 90th percentile stretches to "79 words" and 533 characters or more. The typical description is roughly twice as long as the space likely to display it.
Our working convention is to front-load the meaningful part into the first 150 characters or so and treat anything after that as a bonus that may never be seen. That is a practical choice about truncation, not a rule Google publishes, and it is worth being clear about the difference when someone quotes a character limit at you.
Specific, constrained, and fed the actual page. The most common mistake is asking a model to write a description from the title alone. It will produce something fluent and generic, because a title does not contain enough information to summarise a page accurately.
Give the model the page's real content, the primary query the page targets, the audience, and a hard constraint on length and tone. Then ask for the single most useful thing the page offers, in plain language, without adjectives. A model that has read the page writes a much better description than one guessing from a headline.
The other instruction that improves output immediately is telling it what not to do. Ban the phrases you are tired of seeing, ban questions as openers if every page already opens with one, and ban restating the title. Google's condition that descriptions stay "diverse" is a real constraint, and models default to sameness unless you push against it.
This is the same discipline we apply to any AI writing task on a site, and we went through the wider version of it in our notes on using AI for website copy.
It invents specifics. A model summarising a services page will happily add a number of years in business, a location, or a claim about turnaround time that appears nowhere on the page. The sentence reads perfectly and is quietly false, which is the worst combination for something that will appear in search results under your brand.
It also flattens distinctions. Ask a model to describe nine similar service pages and you will get nine descriptions that could be swapped between pages without anyone noticing. That directly fails Google's diversity condition, and it removes the only reason to have written separate pages in the first place.
The third failure is tone drift. Models reach for superlatives by default, so descriptions arrive full of leading, comprehensive, and cutting-edge unless you explicitly forbid them. On a site that has worked hard on its voice, that is a visible regression even when the facts are right.
None of these are reasons to avoid the tool. They are reasons to review the output as a fact-checking task rather than a proofreading one, which is a distinction we made in our piece on fact-checking AI content.
More cautiously, because titles do influence rankings and they are far less forgiving. A title has to carry the primary keyword, read naturally, and fit a much tighter space. There is less room for a model to be approximately right.
The Web Almanac puts median title length at "12 words" and "77 characters on desktop and 79 on mobile", which is already at the edge of what displays. Ask a model for a title and it will usually overshoot, then pad with your brand name and a separator that eats another ten characters.
Our approach is to let AI propose titles and never accept them unedited. Descriptions can survive a light review at volume. Titles are the single most important on-page signal you control, and spending a minute each on the pages that matter is not the place to optimise for speed.
Review the pattern, not each line. Export every description into a spreadsheet next to its URL and target query, then sort and scan for the three failure modes rather than reading them as prose. Duplicates, invented facts, and repeated openers all surface far faster in a sorted column than in a page-by-page review.
Start by sorting alphabetically, which makes near-duplicates sit next to each other and become obvious immediately. Then scan for numbers, dates, and place names, because those are where fabrication concentrates. Every one of them needs checking against the actual page, and this is the step nobody should skip.
A crawler makes the last part straightforward. Screaming Frog, Ahrefs, or Semrush will all pull every description on the site and flag duplicates and outliers, which turns a reading task into a filtering task. That is the difference between a review that happens and one that gets postponed forever.
Less than for Google, and it is worth being honest about that rather than overselling the connection. Answer engines summarise page content directly, so a meta description is not the thing getting cited. Well-structured page content is.
What the same 2025 Web Almanac data does show is how early most of this still is. Its generative AI chapter found that GPTBot directives in robots.txt "rose from 2.6% in 2024 to around 4.5% in 2025", and that of desktop pages, "only 2.13% exhibit valid llms.txt entries". The infrastructure people argue about online is barely deployed in practice.
Our reading is that the fundamentals still pay better than the novelties. A site with accurate, distinct descriptions on every page and clean structured content will do well in both places. A site chasing AI-specific files while a third of its pages have no description at all has the order backwards.
Find out how many of your pages are missing a description. Run a crawl, filter for empty and duplicate descriptions, and count them. If you are near the 67% adoption figure from the Web Almanac, roughly a third of your pages are showing Google whatever it can find.
Then generate in batches, review by pattern, and fix the highest traffic pages by hand. Let the tool handle the long tail where nobody was ever going to write them anyway, and spend real attention on the twenty pages that bring in most of your visitors. That split is where the value actually is.
We use AI for exactly this kind of work, and we are careful with it because a fluent false sentence under your brand in search results is worse than a blank field. If you want help setting up a process that stays accurate at volume, we are happy to walk through how we do it. You can find us at phoenix.studio.
Tell us where you want to go. We'll tell you how we'd get you there.